From 44b45ab17aec08a8aa6528a5b7568e2ab600eb82 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Mon, 29 Jun 2026 21:37:58 +0200
Subject: [PATCH 01/29] [RNE Rewrite] feat(ocr): unified on-device OCR +
 document pipeline

Two-stage OCR (EasyOCR CRAFT+CRNN / PaddleOCR DBNet+SVTR) plus a document
pipeline, on top of rne-rewrite.

- One fused PTE per model with bucketed detect_<S>/recognize_<W> methods and
  snap-to-closest sizing; a single baked contract, with only the box decoder
  (detectorKind: 'craft' | 'dbnet') and the drop score per architecture.
- Document pipeline: layout via createObjectDetector, native dewarp/gridSample,
  SLANet_plus table-structure recognition, structure-guided table HTML.
- Vertical reading (additive, opt-in): page-level column grouping for stacked
  signage + char-level second CRAFT pass + joint-hconcat recognition; tall lines
  are no longer flipped flat, and vertical reads skip the drop-score gate.
- Native ops: extractTextBoxes (CRAFT + DBNet), warpQuad, ctcGreedyDecode, gridSample.
- Models hosted on Hugging Face (EasyOCR, PP-OCRv6, PP-DocLayoutV3, PaddleHelpers),
  downloaded + cached on device; demo screens consume them directly.
---
 .cspell-wordlist.txt                          |  32 +
 apps/computer-vision/app/_layout.tsx          |  14 +
 apps/computer-vision/app/document/index.tsx   | 383 ++++++++
 apps/computer-vision/app/index.tsx            |   6 +
 apps/computer-vision/app/ocr/index.tsx        | 307 +++++++
 .../components/ImageViewport.tsx              |  57 +-
 .../cpp/extensions/cv/box_ops.cpp             |   2 +
 .../cpp/extensions/cv/image_ops.cpp           | 105 +++
 .../cpp/extensions/cv/image_ops.h             |   1 +
 .../cpp/extensions/cv/install.cpp             |   6 +
 .../cpp/extensions/cv/ocr_ops.cpp             | 826 ++++++++++++++++++
 .../cpp/extensions/cv/ocr_ops.h               |  17 +
 .../react-native-executorch/src/constants.ts  | 102 +++
 .../src/extensions/cv/ops/boxes.ts            |  24 +-
 .../src/extensions/cv/ops/image.ts            | 192 ++++
 .../src/extensions/cv/ops/points.ts           |  13 +
 .../extensions/cv/tasks/documentHelpers.ts    | 179 ++++
 .../src/extensions/cv/tasks/documentOCR.ts    | 214 +++++
 .../src/extensions/cv/tasks/ocr.ts            | 816 +++++++++++++++++
 .../src/extensions/cv/tasks/ocrHelpers.ts     | 263 ++++++
 .../src/extensions/cv/tasks/supporting.ts     | 261 ++++++
 .../src/hooks/useDocumentOCR.ts               |  59 ++
 .../src/hooks/useOCR.ts                       |  39 +
 packages/react-native-executorch/src/index.ts |   7 +
 .../react-native-executorch/src/models.ts     | 134 +++
 .../react-native-executorch/src/ocrSymbols.ts | 161 ++++
 26 files changed, 4211 insertions(+), 9 deletions(-)
 create mode 100644 apps/computer-vision/app/document/index.tsx
 create mode 100644 apps/computer-vision/app/ocr/index.tsx
 create mode 100644 packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
 create mode 100644 packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
 create mode 100644 packages/react-native-executorch/src/hooks/useDocumentOCR.ts
 create mode 100644 packages/react-native-executorch/src/hooks/useOCR.ts
 create mode 100644 packages/react-native-executorch/src/ocrSymbols.ts

diff --git a/.cspell-wordlist.txt b/.cspell-wordlist.txt
index aed3ee1e7d..77307d75eb 100644
--- a/.cspell-wordlist.txt
+++ b/.cspell-wordlist.txt
@@ -275,3 +275,35 @@ pcre
 libkleidicv
 thresholding
 binarization
+dbnet
+svtr
+softmaxed
+softmax
+unclip
+cand
+parameterizes
+pyimagesearch
+letterbox
+CRNN
+CRAFT
+PaddleOCR
+EasyOCR
+cornerMean
+ctc
+Vatti
+softmaxing
+ppocrv
+PPOCRV
+ctcGreedyDecode
+dewarp
+vctx
+onehot
+slanet
+letterboxed
+redetect
+redetections
+eos
+doclayout
+dynint
+softmaxes
+hconcat
diff --git a/apps/computer-vision/app/_layout.tsx b/apps/computer-vision/app/_layout.tsx
index f269b27aae..9f3fadd51b 100644
--- a/apps/computer-vision/app/_layout.tsx
+++ b/apps/computer-vision/app/_layout.tsx
@@ -62,6 +62,20 @@ export default function Layout() {
           title: 'Instance Segmentation',
         }}
       />
+      <Drawer.Screen
+        name="ocr/index"
+        options={{
+          drawerLabel: 'OCR',
+          title: 'OCR',
+        }}
+      />
+      <Drawer.Screen
+        name="document/index"
+        options={{
+          drawerLabel: 'Document Pipeline',
+          title: 'Document Pipeline',
+        }}
+      />
       <Drawer.Screen
         name="inspect/index"
         options={{
diff --git a/apps/computer-vision/app/document/index.tsx b/apps/computer-vision/app/document/index.tsx
new file mode 100644
index 0000000000..3aaf1c070a
--- /dev/null
+++ b/apps/computer-vision/app/document/index.tsx
@@ -0,0 +1,383 @@
+import React, { useState } from 'react';
+import { View, Text, StyleSheet, ScrollView, Switch, Platform } from 'react-native';
+import { commonStyles, ColorPalette } from '../../theme';
+import { useImage, Skia, ColorType, AlphaType, type SkImage } from '@shopify/react-native-skia';
+import { useDocumentOCR, models } from 'react-native-executorch';
+import ScreenWrapper from '../../components/ScreenWrapper';
+import { getImage } from '../../utils';
+import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
+import { ImageViewport } from '../../components/ImageViewport';
+import { ModelStatus } from '../../components/ModelStatus';
+import { Button } from '../../components/Button';
+
+const PREVIEW_HEIGHT = 280;
+
+// Hosted per-backend model triplets (OCR + layout + supporting) — downloaded +
+// cached on-device from Hugging Face by `useDocumentOCR`. Backends are filtered
+// by platform (Vulkan = Android, CoreML = iOS, XNNPACK = both).
+type BackendKey = 'XNNPACK' | 'VULKAN' | 'COREML';
+const BACKENDS: { key: BackendKey; label: string; platforms: string[] }[] = [
+  { key: 'XNNPACK', label: 'XNNPACK (CPU)', platforms: ['ios', 'android'] },
+  { key: 'VULKAN', label: 'Vulkan (GPU)', platforms: ['android'] },
+  { key: 'COREML', label: 'CoreML (ANE)', platforms: ['ios'] },
+];
+const AVAILABLE = BACKENDS.filter((b) => b.platforms.includes(Platform.OS));
+const BACKEND_OPTIONS: ModelOption[] = AVAILABLE.map((b, i) => ({ label: b.label, value: i }));
+
+type Cell = { text: string; colspan: number; rowspan: number };
+type DocBlock = {
+  regionType: string;
+  text: string;
+  isTable: boolean;
+  tableHtml?: string;
+  bbox: { xmin: number; ymin: number; xmax: number; ymax: number };
+};
+
+// Parse the SLANet structure HTML (filled) into rows of cells for rendering.
+function parseTable(html: string): Cell[][] {
+  const rows: Cell[][] = [];
+  const trRe = /<tr>([\s\S]*?)<\/tr>/g;
+  let tr: RegExpExecArray | null;
+  while ((tr = trRe.exec(html))) {
+    const cells: Cell[] = [];
+    const tdRe = /<td([^>]*)>([\s\S]*?)<\/td>/g;
+    let td: RegExpExecArray | null;
+    while ((td = tdRe.exec(tr[1]!))) {
+      const attrs = td[1] ?? '';
+      cells.push({
+        text: td[2] ?? '',
+        colspan: Number(/colspan="(\d+)"/.exec(attrs)?.[1] ?? 1),
+        rowspan: Number(/rowspan="(\d+)"/.exec(attrs)?.[1] ?? 1),
+      });
+    }
+    rows.push(cells);
+  }
+  return rows;
+}
+
+function TableView({ html }: { html: string }) {
+  const rows = parseTable(html);
+  if (rows.length === 0) {
+    return <Text style={styles.blockText}>{html}</Text>;
+  }
+  // Fixed-width cells inside a horizontal scroll — wide tables scroll instead of
+  // squishing every column into the screen width.
+  return (
+    <ScrollView horizontal showsHorizontalScrollIndicator={false}>
+      <View style={styles.table}>
+        {rows.map((cells, r) => (
+          <View key={r} style={styles.tr}>
+            {cells.map((c, i) => (
+              <View key={i} style={[styles.td, { width: 110 * c.colspan }]}>
+                <Text style={styles.tdText}>{c.text}</Text>
+              </View>
+            ))}
+          </View>
+        ))}
+      </View>
+    </ScrollView>
+  );
+}
+
+function DocumentContent() {
+  const [backendIdx, setBackendIdx] = useState(0);
+  const [layoutOn, setLayoutOn] = useState(true);
+  const [supportingOn, setSupportingOn] = useState(true);
+  const [orientation, setOrientation] = useState(true);
+  const [dewarp, setDewarp] = useState(true);
+  const [imageUri, setImageUri] = useState<string | null>(null);
+  const [isProcessing, setIsProcessing] = useState(false);
+  const [blocks, setBlocks] = useState<DocBlock[]>([]);
+  // The frame the result boxes are relative to (orientation/dewarp may move it
+  // away from the original), so the overlay lines up.
+  const [processed, setProcessed] = useState<SkImage | null>(null);
+  const [wallMs, setWallMs] = useState<number | null>(null);
+  const [error, setError] = useState<string | null>(null);
+
+  const backend = AVAILABLE[backendIdx]!;
+
+  const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
+
+  // Hosted configs — `useDocumentOCR` downloads + caches each enabled model.
+  const config = {
+    ocr: models.ocr.PADDLE.PPOCRV6_SMALL[backend.key],
+    ...(layoutOn ? { layout: models.layoutDetection.PP_DOCLAYOUT[backend.key] } : {}),
+    ...(supportingOn ? { supporting: models.supporting.PP_SUPPORTING[backend.key] } : {}),
+    orientation,
+    dewarp,
+  };
+
+  const { isReady, downloadProgress, error: loadError, runDocumentOCR } = useDocumentOCR(config);
+
+  const handlePick = async (useCamera: boolean) => {
+    setError(null);
+    try {
+      const uri = await getImage(useCamera);
+      if (uri) {
+        setImageUri(uri);
+        setBlocks([]);
+        setProcessed(null);
+        setWallMs(null);
+      }
+    } catch (e: any) {
+      setError(e.message || String(e));
+    }
+  };
+
+  const run = async () => {
+    if (!skiaImage || !runDocumentOCR) return;
+    setIsProcessing(true);
+    setError(null);
+    try {
+      const pixels = skiaImage.readPixels();
+      if (!(pixels instanceof Uint8Array)) throw new Error('Expected Uint8Array from readPixels');
+      const start = Date.now();
+      const out = await runDocumentOCR({
+        data: pixels,
+        width: skiaImage.width(),
+        height: skiaImage.height(),
+        format: 'rgba' as const,
+        layout: 'hwc' as const,
+      });
+      setWallMs(Date.now() - start);
+      setBlocks(out.blocks as DocBlock[]);
+      // Show the frame the boxes are relative to (orientation/dewarp may have
+      // rotated/warped it), so the overlaid boxes align.
+      const frame = out.image;
+      const skData = Skia.Data.fromBytes(frame.data);
+      const frameImage = Skia.Image.MakeImage(
+        {
+          width: frame.width,
+          height: frame.height,
+          colorType: ColorType.RGBA_8888,
+          alphaType: AlphaType.Unpremul,
+        },
+        skData,
+        frame.width * 4
+      );
+      setProcessed(frameImage);
+    } catch (e: any) {
+      setError(e.message || String(e));
+    } finally {
+      setIsProcessing(false);
+    }
+  };
+
+  const activeError = loadError ? String(loadError) : error;
+  const boxes = blocks.map((b) => [
+    { x: b.bbox.xmin, y: b.bbox.ymin },
+    { x: b.bbox.xmax, y: b.bbox.ymin },
+    { x: b.bbox.xmax, y: b.bbox.ymax },
+    { x: b.bbox.xmin, y: b.bbox.ymax },
+  ]);
+
+  return (
+    <ScrollView
+      style={commonStyles.container}
+      contentContainerStyle={commonStyles.contentContainer}
+    >
+      <Text style={commonStyles.description}>
+        Full document pipeline: layout → OCR grouped into reading-ordered blocks, with
+        orientation/dewarp and table-structure recognition. PaddleOCR is always on; the rest are on
+        by default — toggle any off (toggling reloads the models).
+      </Text>
+
+      <ModelPicker
+        label="Backend"
+        options={BACKEND_OPTIONS}
+        selectedValue={backendIdx}
+        onValueChange={(v) => {
+          setBackendIdx(v);
+          setBlocks([]);
+          setProcessed(null);
+          setWallMs(null);
+        }}
+      />
+
+      <Toggle label="Layout (blocks)" value={layoutOn} onChange={setLayoutOn} />
+      <Toggle label="Tables" value={supportingOn} onChange={setSupportingOn} />
+      <Toggle
+        label="Correct orientation"
+        value={orientation}
+        onChange={setOrientation}
+        hint="needs Tables on"
+      />
+      <Toggle label="Dewarp" value={dewarp} onChange={setDewarp} hint="needs Tables on" />
+
+      <ModelStatus
+        isReady={isReady}
+        downloadProgress={downloadProgress}
+        error={activeError}
+        modelTypeLabel="document pipeline"
+      />
+
+      <ImageViewport
+        skiaImage={processed ?? skiaImage}
+        height={PREVIEW_HEIGHT}
+        boxes={boxes}
+        onPressPlaceholder={() => handlePick(false)}
+      />
+
+      <View style={commonStyles.buttonRow}>
+        <Button title="Gallery" onPress={() => handlePick(false)} variant="secondary" />
+        <Button title="Camera" onPress={() => handlePick(true)} variant="secondary" />
+      </View>
+      <View style={commonStyles.buttonRow}>
+        <Button
+          title="Run pipeline"
+          onPress={run}
+          disabled={!skiaImage || !isReady || isProcessing}
+          loading={isProcessing}
+        />
+      </View>
+
+      {wallMs !== null && (
+        <View style={styles.statsCard}>
+          <Text style={styles.statsTitle}>Performance</Text>
+          <View style={styles.statTiles}>
+            <View style={styles.tile}>
+              <Text style={styles.tileValue}>
+                {wallMs}
+                <Text style={styles.tileUnit}> ms</Text>
+              </Text>
+              <Text style={styles.tileLabel}>Wall time</Text>
+            </View>
+            <View style={styles.tile}>
+              <Text style={styles.tileValue}>{blocks.length}</Text>
+              <Text style={styles.tileLabel}>Blocks</Text>
+            </View>
+          </View>
+        </View>
+      )}
+
+      {blocks.length > 0 && (
+        <View style={styles.results}>
+          <Text style={styles.resultsTitle}>Blocks ({blocks.length})</Text>
+          {blocks.map((b, i) => (
+            <View key={i} style={styles.block}>
+              <Text style={styles.regionType}>
+                {b.regionType}
+                {b.isTable ? '  · table' : ''}
+              </Text>
+              {b.isTable && b.tableHtml ? (
+                <TableView html={b.tableHtml} />
+              ) : (
+                <Text style={styles.blockText}>{b.text}</Text>
+              )}
+            </View>
+          ))}
+        </View>
+      )}
+    </ScrollView>
+  );
+}
+
+function Toggle({
+  label,
+  value,
+  onChange,
+  hint,
+}: {
+  label: string;
+  value: boolean;
+  onChange: (v: boolean) => void;
+  hint?: string;
+}) {
+  return (
+    <View style={styles.toggleRow}>
+      <View style={styles.toggleText}>
+        <Text style={styles.toggleLabel}>{label}</Text>
+        {hint ? <Text style={styles.toggleHint}>{hint}</Text> : null}
+      </View>
+      <Switch value={value} onValueChange={onChange} />
+    </View>
+  );
+}
+
+export default function DocumentScreen() {
+  return (
+    <ScreenWrapper>
+      <DocumentContent />
+    </ScreenWrapper>
+  );
+}
+
+const styles = StyleSheet.create({
+  toggleRow: {
+    flexDirection: 'row',
+    alignItems: 'center',
+    justifyContent: 'space-between',
+    width: '100%',
+    marginBottom: 8,
+  },
+  toggleText: { flex: 1, marginRight: 12 },
+  toggleLabel: { fontSize: 15, fontWeight: '600', color: ColorPalette.strongPrimary },
+  toggleHint: { fontSize: 12, color: '#868e96', marginTop: 2 },
+  statsCard: {
+    width: '100%',
+    backgroundColor: '#fff',
+    borderRadius: 12,
+    padding: 16,
+    marginBottom: 16,
+    borderWidth: 1,
+    borderColor: '#e9ecef',
+  },
+  statsTitle: {
+    fontSize: 12,
+    fontWeight: '700',
+    letterSpacing: 1,
+    color: '#868e96',
+    textTransform: 'uppercase',
+    marginBottom: 12,
+  },
+  statTiles: { flexDirection: 'row', gap: 12 },
+  tile: {
+    flex: 1,
+    backgroundColor: '#f2f4ff',
+    borderRadius: 10,
+    paddingVertical: 12,
+    paddingHorizontal: 14,
+  },
+  tileValue: {
+    fontSize: 24,
+    fontWeight: '800',
+    color: '#001A72',
+    fontVariant: ['tabular-nums'],
+  },
+  tileUnit: { fontSize: 14, fontWeight: '600', color: '#6b73a3' },
+  tileLabel: { fontSize: 11, color: '#868e96', marginTop: 4 },
+  results: {
+    width: '100%',
+    backgroundColor: '#fff',
+    borderRadius: 12,
+    padding: 16,
+    borderWidth: 1,
+    borderColor: '#e9ecef',
+    marginTop: 12,
+  },
+  resultsTitle: {
+    fontSize: 16,
+    fontWeight: '600',
+    color: ColorPalette.strongPrimary,
+    marginBottom: 12,
+  },
+  block: { paddingVertical: 8, borderBottomWidth: 1, borderBottomColor: '#f1f3f5' },
+  regionType: {
+    fontSize: 12,
+    fontWeight: '700',
+    color: '#2b8a3e',
+    textTransform: 'uppercase',
+    marginBottom: 4,
+  },
+  blockText: { fontSize: 14, color: '#333' },
+  table: { borderWidth: 1, borderColor: '#ced4da', borderRadius: 4, overflow: 'hidden' },
+  tr: { flexDirection: 'row' },
+  td: {
+    borderWidth: StyleSheet.hairlineWidth,
+    borderColor: '#ced4da',
+    paddingHorizontal: 6,
+    paddingVertical: 4,
+    minWidth: 24,
+  },
+  tdText: { fontSize: 13, color: '#333' },
+});
diff --git a/apps/computer-vision/app/index.tsx b/apps/computer-vision/app/index.tsx
index cc44604e67..291450a2a8 100644
--- a/apps/computer-vision/app/index.tsx
+++ b/apps/computer-vision/app/index.tsx
@@ -32,6 +32,12 @@ export default function Home() {
         <TouchableOpacity style={styles.button} onPress={() => router.navigate('keypoint/')}>
           <Text style={styles.buttonText}>Keypoint Detection</Text>
         </TouchableOpacity>
+        <TouchableOpacity style={styles.button} onPress={() => router.navigate('ocr/')}>
+          <Text style={styles.buttonText}>OCR</Text>
+        </TouchableOpacity>
+        <TouchableOpacity style={styles.button} onPress={() => router.navigate('document/')}>
+          <Text style={styles.buttonText}>Document Pipeline</Text>
+        </TouchableOpacity>
         <TouchableOpacity style={styles.button} onPress={() => router.navigate('inspect/')}>
           <Text style={styles.buttonText}>Model Inspector</Text>
         </TouchableOpacity>
diff --git a/apps/computer-vision/app/ocr/index.tsx b/apps/computer-vision/app/ocr/index.tsx
new file mode 100644
index 0000000000..a0c3788521
--- /dev/null
+++ b/apps/computer-vision/app/ocr/index.tsx
@@ -0,0 +1,307 @@
+import React, { useState } from 'react';
+import { View, Text, StyleSheet, ScrollView, Platform, Switch } from 'react-native';
+import { commonStyles, ColorPalette } from '../../theme';
+import { useImage } from '@shopify/react-native-skia';
+import { useOCR, models, type OCRDetection } from 'react-native-executorch';
+import ScreenWrapper from '../../components/ScreenWrapper';
+import { getImage } from '../../utils';
+import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
+import { ImageViewport } from '../../components/ImageViewport';
+import { ModelStatus } from '../../components/ModelStatus';
+import { Button } from '../../components/Button';
+
+const PREVIEW_HEIGHT = 280;
+
+// Hosted PTEs — downloaded + cached on-device from Hugging Face by `useOCR`.
+// Backends per platform: XNNPACK runs everywhere, Vulkan on Android, CoreML on iOS.
+const ALL_MODELS = [
+  {
+    label: 'PaddleOCR (XNNPACK)',
+    config: models.ocr.PADDLE.PPOCRV6_SMALL.XNNPACK,
+    platforms: ['ios', 'android'],
+  },
+  {
+    label: 'PaddleOCR (Vulkan)',
+    config: models.ocr.PADDLE.PPOCRV6_SMALL.VULKAN,
+    platforms: ['android'],
+  },
+  {
+    label: 'PaddleOCR (CoreML)',
+    config: models.ocr.PADDLE.PPOCRV6_SMALL.COREML,
+    platforms: ['ios'],
+  },
+  {
+    label: 'EasyOCR English (XNNPACK)',
+    config: models.ocr.EASYOCR.ENGLISH.XNNPACK,
+    platforms: ['ios', 'android'],
+  },
+  {
+    label: 'EasyOCR English (Vulkan)',
+    config: models.ocr.EASYOCR.ENGLISH.VULKAN,
+    platforms: ['android'],
+  },
+  {
+    label: 'EasyOCR English (CoreML)',
+    config: models.ocr.EASYOCR.ENGLISH.COREML,
+    platforms: ['ios'],
+  },
+];
+
+const OCR_MODELS = ALL_MODELS.filter((m) => m.platforms.includes(Platform.OS));
+
+const MODEL_OPTIONS: ModelOption[] = OCR_MODELS.map((m, i) => ({ label: m.label, value: i }));
+
+function OCRContent() {
+  const [selectedIdx, setSelectedIdx] = useState(0);
+  const [vertical, setVertical] = useState(false);
+  const [imageUri, setImageUri] = useState<string | null>(null);
+  const [isProcessing, setIsProcessing] = useState(false);
+  const [results, setResults] = useState<OCRDetection[]>([]);
+  const [wallMs, setWallMs] = useState<number | null>(null);
+  const [error, setError] = useState<string | null>(null);
+
+  const selected = OCR_MODELS[selectedIdx]!;
+
+  const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
+
+  // `useOCR` downloads + caches the hosted PTE from its Hugging Face URL.
+  const { isReady, downloadProgress, error: loadError, runOCR } = useOCR(selected.config);
+
+  const handlePickImage = async (useCamera: boolean) => {
+    setError(null);
+    try {
+      const uri = await getImage(useCamera);
+      if (uri) {
+        setImageUri(uri);
+        setResults([]);
+        setWallMs(null);
+      }
+    } catch (e: any) {
+      setError(e.message || String(e));
+    }
+  };
+
+  const runRecognition = async () => {
+    if (!skiaImage || !runOCR) return;
+    setIsProcessing(true);
+    setError(null);
+    try {
+      const pixels = skiaImage.readPixels();
+      if (!(pixels instanceof Uint8Array)) {
+        throw new Error('Expected Uint8Array from readPixels');
+      }
+      const buffer = {
+        data: pixels,
+        width: skiaImage.width(),
+        height: skiaImage.height(),
+        format: 'rgba' as const,
+        layout: 'hwc' as const,
+      };
+      const start = Date.now();
+      // `vertical` is a per-run option now — toggling it needs no model reload.
+      const output = await runOCR(buffer, { vertical });
+      setWallMs(Date.now() - start);
+      setResults(output.detections);
+    } catch (e: any) {
+      setError(e.message || String(e));
+    } finally {
+      setIsProcessing(false);
+    }
+  };
+
+  const activeError = loadError ? String(loadError) : error;
+
+  return (
+    <ScrollView
+      style={commonStyles.container}
+      contentContainerStyle={commonStyles.contentContainer}
+    >
+      <Text style={commonStyles.description}>
+        Upload or capture an image to detect and recognize text on-device.
+      </Text>
+
+      <ModelPicker
+        label="Model"
+        options={MODEL_OPTIONS}
+        selectedValue={selectedIdx}
+        onValueChange={(idx) => {
+          setSelectedIdx(idx);
+          setResults([]);
+          setWallMs(null);
+          setError(null);
+        }}
+      />
+
+      <View style={styles.toggleRow}>
+        <View style={styles.toggleText}>
+          <Text style={styles.toggleLabel}>Vertical text</Text>
+          <Text style={styles.toggleHint}>
+            Read upright stacked columns (character-under-character)
+          </Text>
+        </View>
+        <Switch value={vertical} onValueChange={setVertical} />
+      </View>
+
+      <ModelStatus
+        isReady={isReady}
+        downloadProgress={downloadProgress}
+        error={activeError}
+        modelTypeLabel="OCR model"
+      />
+
+      <ImageViewport
+        skiaImage={skiaImage}
+        height={PREVIEW_HEIGHT}
+        boxes={results.map((r) => r.quad)}
+        onPressPlaceholder={() => handlePickImage(false)}
+      />
+
+      <View style={commonStyles.buttonRow}>
+        <Button title="Gallery" onPress={() => handlePickImage(false)} variant="secondary" />
+        <Button title="Camera" onPress={() => handlePickImage(true)} variant="secondary" />
+      </View>
+
+      <View style={commonStyles.buttonRow}>
+        <Button
+          title="Run OCR"
+          onPress={runRecognition}
+          disabled={!skiaImage || !isReady || isProcessing}
+          loading={isProcessing}
+        />
+      </View>
+
+      {wallMs !== null && (
+        <View style={styles.statsCard}>
+          <Text style={styles.statsTitle}>Performance</Text>
+          <View style={styles.statTiles}>
+            <View style={styles.tile}>
+              <Text style={styles.tileValue}>
+                {wallMs}
+                <Text style={styles.tileUnit}> ms</Text>
+              </Text>
+              <Text style={styles.tileLabel}>Wall time</Text>
+            </View>
+            <View style={styles.tile}>
+              <Text style={styles.tileValue}>{results.length}</Text>
+              <Text style={styles.tileLabel}>Regions read</Text>
+            </View>
+          </View>
+        </View>
+      )}
+
+      {results.length > 0 && (
+        <View style={styles.resultsContainer}>
+          <Text style={styles.resultsTitle}>Detected text ({results.length})</Text>
+          {results.map((res, idx) => (
+            <View key={idx} style={styles.resultRow}>
+              <Text style={styles.resultLabel} numberOfLines={1}>
+                {res.text}
+              </Text>
+              <View style={styles.resultMeta}>
+                <Text style={styles.resultMs}>{res.recognizeMs.toFixed(0)}ms</Text>
+                <Text style={styles.resultConfidence}>{Math.round(res.confidence * 100)}%</Text>
+              </View>
+            </View>
+          ))}
+        </View>
+      )}
+    </ScrollView>
+  );
+}
+
+export default function OCRScreen() {
+  return (
+    <ScreenWrapper>
+      <OCRContent />
+    </ScreenWrapper>
+  );
+}
+
+const styles = StyleSheet.create({
+  toggleRow: {
+    flexDirection: 'row',
+    alignItems: 'center',
+    justifyContent: 'space-between',
+    width: '100%',
+    marginBottom: 12,
+  },
+  toggleText: { flex: 1, marginRight: 12 },
+  toggleLabel: { fontSize: 15, fontWeight: '600', color: ColorPalette.strongPrimary },
+  toggleHint: { fontSize: 12, color: '#868e96', marginTop: 2 },
+  statsCard: {
+    width: '100%',
+    backgroundColor: '#fff',
+    borderRadius: 12,
+    padding: 16,
+    marginBottom: 16,
+    borderWidth: 1,
+    borderColor: '#e9ecef',
+  },
+  statsTitle: {
+    fontSize: 12,
+    fontWeight: '700',
+    letterSpacing: 1,
+    color: '#868e96',
+    textTransform: 'uppercase',
+    marginBottom: 12,
+  },
+  statTiles: {
+    flexDirection: 'row',
+    gap: 12,
+    marginBottom: 12,
+  },
+  tile: {
+    flex: 1,
+    backgroundColor: '#f2f4ff',
+    borderRadius: 10,
+    paddingVertical: 12,
+    paddingHorizontal: 14,
+  },
+  tileValue: {
+    fontSize: 24,
+    fontWeight: '800',
+    color: '#001A72',
+    fontVariant: ['tabular-nums'],
+  },
+  tileUnit: { fontSize: 14, fontWeight: '600', color: '#6b73a3' },
+  tileLabel: { fontSize: 11, color: '#868e96', marginTop: 4 },
+  resultMeta: { flexDirection: 'row', alignItems: 'center' },
+  resultMs: {
+    fontSize: 12,
+    color: '#868e96',
+    marginRight: 10,
+    fontVariant: ['tabular-nums'],
+  },
+  resultsContainer: {
+    width: '100%',
+    backgroundColor: '#fff',
+    borderRadius: 12,
+    padding: 16,
+    borderWidth: 1,
+    borderColor: '#e9ecef',
+  },
+  resultsTitle: {
+    fontSize: 16,
+    fontWeight: '600',
+    color: ColorPalette.strongPrimary,
+    marginBottom: 12,
+  },
+  resultRow: {
+    flexDirection: 'row',
+    justifyContent: 'space-between',
+    paddingVertical: 8,
+    borderBottomWidth: 1,
+    borderBottomColor: '#f1f3f5',
+  },
+  resultLabel: {
+    fontSize: 14,
+    color: '#333',
+    flex: 1,
+    marginRight: 8,
+  },
+  resultConfidence: {
+    fontSize: 14,
+    fontWeight: '600',
+    color: '#2b8a3e',
+  },
+});
diff --git a/apps/computer-vision/components/ImageViewport.tsx b/apps/computer-vision/components/ImageViewport.tsx
index 593133eaf6..701c83d655 100644
--- a/apps/computer-vision/components/ImageViewport.tsx
+++ b/apps/computer-vision/components/ImageViewport.tsx
@@ -1,16 +1,23 @@
-import React from 'react';
+import React, { useMemo } from 'react';
 import { View, Text, StyleSheet, TouchableOpacity, Dimensions } from 'react-native';
 import {
   Canvas,
   Image as SkImage,
   BlendColor,
+  Path,
+  Skia,
   type SkImage as SkiaImageType,
 } from '@shopify/react-native-skia';
 
 import { theme } from '../theme';
 
 const VIEW_WIDTH = Dimensions.get('window').width - 32;
-const VIEW_HEIGHT = Math.round((VIEW_WIDTH * 16) / 9);
+const DEFAULT_VIEW_HEIGHT = Math.round((VIEW_WIDTH * 16) / 9);
+
+/** A 2D point in original-image pixel coordinates. */
+type Point = { readonly x: number; readonly y: number };
+/** A polygon (e.g. an OCR quad) in original-image pixel coordinates. */
+type Polygon = readonly Point[];
 
 export interface ImageViewportProps {
   skiaImage: SkiaImageType | null;
@@ -20,6 +27,10 @@ export interface ImageViewportProps {
   placeholderText?: string;
   overlayOpacity?: number;
   children?: React.ReactNode;
+  /** Height of the preview box in px. Defaults to a 16:9 box. */
+  height?: number;
+  /** Polygons (in original image px) to stroke over the image, e.g. OCR quads. */
+  boxes?: readonly Polygon[];
 }
 
 export function ImageViewport({
@@ -30,17 +41,47 @@ export function ImageViewport({
   placeholderText = 'Tap to select an image from gallery',
   overlayOpacity = 0.8,
   children,
+  height,
+  boxes,
 }: ImageViewportProps) {
+  const viewHeight = height ?? DEFAULT_VIEW_HEIGHT;
+
+  // Map original-pixel polygons into canvas space using the same contain-fit
+  // transform Skia uses to draw the image, then build one stroked path.
+  const boxesPath = useMemo(() => {
+    if (!skiaImage || !boxes?.length) return null;
+    const ow = skiaImage.width();
+    const oh = skiaImage.height();
+    if (ow === 0 || oh === 0) return null;
+    const scale = Math.min(VIEW_WIDTH / ow, viewHeight / oh);
+    const dx = (VIEW_WIDTH - ow * scale) / 2;
+    const dy = (viewHeight - oh * scale) / 2;
+
+    const path = Skia.Path.Make();
+    for (const poly of boxes) {
+      if (poly.length < 2) continue;
+      path.moveTo(dx + poly[0]!.x * scale, dy + poly[0]!.y * scale);
+      for (let i = 1; i < poly.length; i++) {
+        path.lineTo(dx + poly[i]!.x * scale, dy + poly[i]!.y * scale);
+      }
+      path.close();
+    }
+    return path;
+  }, [skiaImage, boxes, viewHeight]);
+
   if (!skiaImage) {
     return (
-      <TouchableOpacity style={styles.placeholder} onPress={onPressPlaceholder}>
+      <TouchableOpacity
+        style={[styles.placeholder, { height: viewHeight }]}
+        onPress={onPressPlaceholder}
+      >
         <Text style={styles.placeholderText}>{placeholderText}</Text>
       </TouchableOpacity>
     );
   }
 
   return (
-    <View style={[styles.canvasWrapper, { width: VIEW_WIDTH, height: VIEW_HEIGHT }]}>
+    <View style={[styles.canvasWrapper, { width: VIEW_WIDTH, height: viewHeight }]}>
       <Canvas style={styles.canvas}>
         <SkImage
           image={skiaImage}
@@ -48,7 +89,7 @@ export function ImageViewport({
           x={0}
           y={0}
           width={VIEW_WIDTH}
-          height={VIEW_HEIGHT}
+          height={viewHeight}
         />
         {overlayImage && (
           <SkImage
@@ -57,7 +98,7 @@ export function ImageViewport({
             x={0}
             y={0}
             width={VIEW_WIDTH}
-            height={VIEW_HEIGHT}
+            height={viewHeight}
             opacity={overlayOpacity}
           />
         )}
@@ -70,12 +111,13 @@ export function ImageViewport({
               x={0}
               y={0}
               width={VIEW_WIDTH}
-              height={VIEW_HEIGHT}
+              height={viewHeight}
               opacity={overlayOpacity}
             >
               <BlendColor color={item.color} mode="srcIn" />
             </SkImage>
           ))}
+        {boxesPath && <Path path={boxesPath} style="stroke" strokeWidth={2} color="#39FF14" />}
       </Canvas>
       {children}
     </View>
@@ -85,7 +127,6 @@ export function ImageViewport({
 const styles = StyleSheet.create({
   placeholder: {
     width: '100%',
-    height: VIEW_HEIGHT,
     borderWidth: 2,
     borderColor: theme.colors.border,
     borderStyle: 'dashed',
diff --git a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
index 501b9245c1..e1de7fccf6 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
@@ -65,6 +65,7 @@ std::array<float, 4> decodeToXyxy(
     case BoxFormat::CXCYWH:
         return {a - c / 2.0f, b - d / 2.0f, a + c / 2.0f, b + d / 2.0f};
     }
+    throw std::invalid_argument("decodeToXyxy: unhandled box format");
 }
 } // namespace
 
@@ -242,6 +243,7 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
             return resultGroups;
         }
         }
+        throw jsi::JSError(rt, "nms: unhandled nmsType");
     };
 
     module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index 13bf66cd81..d9b0d26404 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -714,4 +714,109 @@ void install_applyColormap(jsi::Runtime &rt, jsi::Object &module) {
     };
     module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
 }
+
+void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
+    auto name = "gridSample";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
+        if (count != 3) {
+            throw jsi::JSError(rt, "Usage: gridSample(src, grid, dst)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[2].isObject() || !args[2].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "gridSample: src, grid, and dst must be Tensors");
+        }
+
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto grid = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[2].asObject(rt).getHostObject<TensorHostObject>(rt);
+
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "gridSample: In-place operations (src == dst) are not supported.");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "gridSample: src and dst must be uint8");
+        }
+        if (grid->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "gridSample: grid must be float32");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "gridSample: src and dst must be [H, W, C]");
+        }
+        if (src->shape_ != dst->shape_) {
+            throw jsi::JSError(rt, "gridSample: src and dst must have the same shape");
+        }
+        // grid is the torch grid_sample field [..,2,gH,gW], channel 0 = x, 1 = y,
+        // normalized to [-1,1] with align_corners=true.
+        const auto &gs = grid->shape_;
+        if (gs.size() < 3 || gs[gs.size() - 3] != 2) {
+            throw jsi::JSError(rt, "gridSample: grid must be [..,2,gH,gW]");
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        std::shared_lock<std::shared_mutex> gridLock(grid->mutex_, std::try_to_lock);
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock() || !gridLock.owns_lock() || !dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "gridSample: a tensor is currently in use");
+        }
+        if (!src->data_ || !grid->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "gridSample: a tensor has been disposed");
+        }
+
+        const int32_t h = src->shape_[0];
+        const int32_t w = src->shape_[1];
+        const int32_t channels = src->shape_[2];
+        const int32_t gridH = gs[gs.size() - 2];
+        const int32_t gridW = gs[gs.size() - 1];
+        const int32_t plane = gridH * gridW;
+        const float *g = reinterpret_cast<const float *>(grid->data_.get());
+
+        // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
+        auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
+            const int32_t x0 = std::clamp(static_cast<int32_t>(std::floor(gx)), 0, gridW - 1);
+            const int32_t y0 = std::clamp(static_cast<int32_t>(std::floor(gy)), 0, gridH - 1);
+            const int32_t x1 = std::min(x0 + 1, gridW - 1);
+            const int32_t y1 = std::min(y0 + 1, gridH - 1);
+            const float dx = gx - static_cast<float>(x0);
+            const float dy = gy - static_cast<float>(y0);
+            const int32_t base = c * plane;
+            const float top = g[base + y0 * gridW + x0] +
+                              (g[base + y0 * gridW + x1] - g[base + y0 * gridW + x0]) * dx;
+            const float bot = g[base + y1 * gridW + x0] +
+                              (g[base + y1 * gridW + x1] - g[base + y1 * gridW + x0]) * dx;
+            return top + (bot - top) * dy;
+        };
+
+        ::cv::Mat mapX(h, w, CV_32F);
+        ::cv::Mat mapY(h, w, CV_32F);
+        for (int32_t oy = 0; oy < h; ++oy) {
+            const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
+                                         static_cast<float>(gridH - 1)
+                                   : 0.0f;
+            float *rowX = mapX.ptr<float>(oy);
+            float *rowY = mapY.ptr<float>(oy);
+            for (int32_t ox = 0; ox < w; ++ox) {
+                const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
+                                             static_cast<float>(gridW - 1)
+                                       : 0.0f;
+                const float nx = sampleGrid(0, gx, gy); // [-1,1]
+                const float ny = sampleGrid(1, gx, gy);
+                rowX[ox] = ((nx + 1.0f) / 2.0f) * static_cast<float>(w - 1);
+                rowY[ox] = ((ny + 1.0f) / 2.0f) * static_cast<float>(h - 1);
+            }
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(h, w, cvType, src->data_.get());
+        ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
+        try {
+            ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("gridSample: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[2]);
+    };
+    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
+}
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
index 893c0957e5..dc24873a66 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
@@ -9,4 +9,5 @@ void install_toChannelsFirst(facebook::jsi::Runtime &rt, facebook::jsi::Object &
 void install_toChannelsLast(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_normalize(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_applyColormap(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+void install_gridSample(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/install.cpp b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
index 0540f45c9b..9fd35bdc22 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/install.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
@@ -1,6 +1,7 @@
 #include "install.h"
 #include "box_ops.h"
 #include "image_ops.h"
+#include "ocr_ops.h"
 
 namespace rnexecutorch::extensions::cv {
 namespace jsi = facebook::jsi;
@@ -14,10 +15,15 @@ void install(facebook::jsi::Runtime &rt, facebook::jsi::Object &module) {
     image_ops::install_toChannelsLast(rt, cvModule);
     image_ops::install_normalize(rt, cvModule);
     image_ops::install_applyColormap(rt, cvModule);
+    image_ops::install_gridSample(rt, cvModule);
 
     box_ops::install_nms(rt, cvModule);
     box_ops::install_restrictToBox(rt, cvModule);
 
+    ocr_ops::install_extractTextBoxes(rt, cvModule);
+    ocr_ops::install_warpQuad(rt, cvModule);
+    ocr_ops::install_ctcGreedyDecode(rt, cvModule);
+
     module.setProperty(rt, "cv", cvModule);
 }
 } // namespace rnexecutorch::extensions::cv
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
new file mode 100644
index 0000000000..19ab24e6ce
--- /dev/null
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -0,0 +1,826 @@
+#include "ocr_ops.h"
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstddef>
+#include <limits>
+#include <numeric>
+#include <optional>
+#include <span>
+#include <stdexcept>
+#include <tuple>
+#include <unordered_set>
+#include <utility>
+#include <vector>
+
+#include <opencv2/imgproc.hpp>
+
+#include "core/dtype.h"
+#include "core/tensor.h"
+
+// Detector postprocessing geometry: CRAFT text-map grouping + DBNet prob-map ->
+// oriented quads. Pure OpenCV, kept native. ctcGreedyDecode (per-timestep argmax
+// + max prob) is native too; the CTC blank-collapse, charset mapping, and
+// confidence aggregation stay in TypeScript.
+namespace rnexecutorch::extensions::cv::ocr_ops {
+namespace jsi = facebook::jsi;
+using TensorHostObject = rnexecutorch::core::tensor::TensorHostObject;
+
+namespace {
+// ----------------------------- geometry types ------------------------------
+struct Box {
+    float x0, y0, x1, y1; // axis-aligned (p1=min, p2=max)
+    float angle = 0.0f;
+    float width() const { return x1 - x0; }
+    float height() const { return y1 - y0; }
+};
+
+struct Quad {
+    std::array<::cv::Point2f, 4> pts;
+    float score = 1.0f;
+    float angle = 0.0f;
+};
+
+float dist(const ::cv::Point2f &a, const ::cv::Point2f &b) {
+    return std::hypot(b.x - a.x, b.y - a.y);
+}
+::cv::Point2f center(const Box &b) {
+    return {(b.x0 + b.x1) * 0.5f, (b.y0 + b.y1) * 0.5f};
+}
+float minSide(const Box &b) { return std::min(b.width(), b.height()); }
+float maxSide(const Box &b) { return std::max(b.width(), b.height()); }
+bool isClose(float a, float b, float eps = 1e-3f) { return std::fabs(a - b) < eps; }
+
+std::array<::cv::Point2f, 4> corners(const Box &b) {
+    return {::cv::Point2f{b.x0, b.y0}, {b.x1, b.y0}, {b.x1, b.y1}, {b.x0, b.y1}};
+}
+
+::cv::Point2f rotateAround(const ::cv::Point2f &p, const ::cv::Point2f &ctr, float rad) {
+    const float tx = p.x - ctr.x;
+    const float ty = p.y - ctr.y;
+    return {tx * std::cos(rad) - ty * std::sin(rad) + ctr.x,
+            tx * std::sin(rad) + ty * std::cos(rad) + ctr.y};
+}
+
+// ------------------------------ CRAFT branch -------------------------------
+std::pair<::cv::Mat, ::cv::Mat> interleavedToMats(std::span<const float> data, ::cv::Size size) {
+    ::cv::Mat textMap(size, CV_32F);
+    ::cv::Mat affinityMap(size, CV_32F);
+    const auto w = static_cast<std::size_t>(size.width);
+    for (std::size_t i = 0; i < data.size(); ++i) {
+        const int32_t x = static_cast<int32_t>((i / 2) % w);
+        const int32_t y = static_cast<int32_t>((i / 2) / w);
+        if (i % 2 == 0) {
+            textMap.at<float>(y, x) = data[i];
+        } else {
+            affinityMap.at<float>(y, x) = data[i];
+        }
+    }
+    return {textMap, affinityMap};
+}
+
+void dilateComponent(::cv::Mat &segMap, const ::cv::Mat &stats, int32_t i, int32_t area,
+                     int32_t imgW, int32_t imgH) {
+    const int32_t x = stats.at<int32_t>(i, ::cv::CC_STAT_LEFT);
+    const int32_t y = stats.at<int32_t>(i, ::cv::CC_STAT_TOP);
+    const int32_t w = stats.at<int32_t>(i, ::cv::CC_STAT_WIDTH);
+    const int32_t h = stats.at<int32_t>(i, ::cv::CC_STAT_HEIGHT);
+    const int32_t dilationRadius =
+        static_cast<int32_t>(std::sqrt(static_cast<double>(area) / std::max(w, h)) * 2);
+    const int32_t sx = std::max(x - dilationRadius, 0);
+    const int32_t ex = std::min(x + w + dilationRadius, imgW);
+    const int32_t sy = std::max(y - dilationRadius, 0);
+    const int32_t ey = std::min(y + h + dilationRadius, imgH);
+    const int32_t kSize = 1 + dilationRadius;
+    ::cv::Mat kernel = ::cv::getStructuringElement(::cv::MORPH_RECT, ::cv::Size(kSize, kSize));
+    ::cv::Mat roi = segMap(::cv::Rect(sx, sy, ex - sx, ey - sy));
+    ::cv::dilate(roi, roi, kernel, ::cv::Point(-1, -1), 1);
+}
+
+std::optional<Box> boxFromComponent(const ::cv::Mat &textMap, const ::cv::Mat &labels,
+                                    const ::cv::Mat &stats, int32_t i, int32_t imgW, int32_t imgH,
+                                    float lowTextThreshold) {
+    const int32_t area = stats.at<int32_t>(i, ::cv::CC_STAT_AREA);
+    if (area < 10) {
+        return std::nullopt;
+    }
+    ::cv::Mat mask = (labels == i);
+    double maxVal;
+    ::cv::minMaxLoc(textMap, nullptr, &maxVal, nullptr, nullptr, mask);
+    if (maxVal < static_cast<double>(lowTextThreshold)) {
+        return std::nullopt;
+    }
+    ::cv::Mat segMap = ::cv::Mat::zeros(textMap.size(), CV_8U);
+    segMap.setTo(255, mask);
+    dilateComponent(segMap, stats, i, area, imgW, imgH);
+
+    std::vector<std::vector<::cv::Point>> contours;
+    ::cv::findContours(segMap, contours, ::cv::RETR_EXTERNAL, ::cv::CHAIN_APPROX_SIMPLE);
+    if (contours.empty()) {
+        return std::nullopt;
+    }
+    ::cv::RotatedRect rr = ::cv::minAreaRect(contours[0]);
+    std::array<::cv::Point2f, 4> v;
+    rr.points(v.data());
+    Box box;
+    box.x0 = std::min({v[0].x, v[1].x, v[2].x, v[3].x});
+    box.y0 = std::min({v[0].y, v[1].y, v[2].y, v[3].y});
+    box.x1 = std::max({v[0].x, v[1].x, v[2].x, v[3].x});
+    box.y1 = std::max({v[0].y, v[1].y, v[2].y, v[3].y});
+    box.angle = rr.angle;
+    return box;
+}
+
+std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityMap,
+                                        float textThreshold, float linkThreshold,
+                                        float lowTextThreshold) {
+    const int32_t imgH = textMap.rows;
+    const int32_t imgW = textMap.cols;
+    ::cv::Mat textScore, affinityScore;
+    ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
+    ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
+                    ::cv::THRESH_BINARY);
+    ::cv::Mat comb = textScore + affinityScore;
+    ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_BINARY);
+    ::cv::Mat binary;
+    comb.convertTo(binary, CV_8UC1);
+
+    ::cv::Mat labels, stats, centroids;
+    const int32_t nLabels = ::cv::connectedComponentsWithStats(binary, labels, stats, centroids, 4);
+
+    std::vector<Box> boxes;
+    boxes.reserve(static_cast<std::size_t>(nLabels));
+    for (int32_t i = 1; i < nLabels; ++i) {
+        auto box = boxFromComponent(textMap, labels, stats, i, imgW, imgH, lowTextThreshold);
+        if (box) {
+            boxes.push_back(*box);
+        }
+    }
+    return boxes;
+}
+
+// fit a line to the two shortest sides' midpoints; returns slope, intercept, vertical?
+std::tuple<float, float, bool> fitLineToShortestSides(const Box &b, float verticalThreshold) {
+    const auto pts = corners(b);
+    std::array<std::pair<float, int>, 4> sides;
+    std::array<::cv::Point2f, 4> mids;
+    for (int i = 0; i < 4; ++i) {
+        const auto &p1 = pts[static_cast<std::size_t>(i)];
+        const auto &p2 = pts[static_cast<std::size_t>((i + 1) % 4)];
+        sides[static_cast<std::size_t>(i)] = {dist(p1, p2), i};
+        mids[static_cast<std::size_t>(i)] = {(p1.x + p2.x) * 0.5f, (p1.y + p2.y) * 0.5f};
+    }
+    std::sort(sides.begin(), sides.end());
+    ::cv::Point2f m1 = mids[static_cast<std::size_t>(sides[0].second)];
+    ::cv::Point2f m2 = mids[static_cast<std::size_t>(sides[1].second)];
+    const bool isVertical = std::fabs(m2.x - m1.x) < verticalThreshold;
+    std::vector<::cv::Point2f> fitPts = {m1, m2};
+    if (isVertical) {
+        for (auto &p : fitPts) {
+            std::swap(p.x, p.y);
+        }
+    }
+    ::cv::Vec4f line;
+    ::cv::fitLine(fitPts, line, ::cv::DIST_L2, 0, 0.01, 0.01);
+    const float m = line[1] / line[0];
+    const float c = line[3] - m * line[2];
+    return {m, c, isVertical};
+}
+
+Box rotateBox(const Box &b, float angleDeg) {
+    const ::cv::Point2f ctr = center(b);
+    const float rad = angleDeg * static_cast<float>(M_PI) / 180.0f;
+    float minX = std::numeric_limits<float>::max();
+    float minY = std::numeric_limits<float>::max();
+    float maxX = std::numeric_limits<float>::lowest();
+    float maxY = std::numeric_limits<float>::lowest();
+    for (const auto &p : corners(b)) {
+        const ::cv::Point2f r = rotateAround(p, ctr, rad);
+        minX = std::min(minX, r.x);
+        minY = std::min(minY, r.y);
+        maxX = std::max(maxX, r.x);
+        maxY = std::max(maxY, r.y);
+    }
+    return {minX, minY, maxX, maxY, b.angle};
+}
+
+float minDistanceBetween(const Box &a, const Box &b) {
+    float md = std::numeric_limits<float>::max();
+    for (const auto &c1 : corners(a)) {
+        for (const auto &c2 : corners(b)) {
+            md = std::min(md, dist(c1, c2));
+        }
+    }
+    return md;
+}
+
+std::optional<std::pair<std::size_t, float>>
+findClosestBox(const std::vector<Box> &boxes, const std::unordered_set<std::size_t> &ignored,
+               const Box &current, bool isVertical, float m, float c, float centerThreshold) {
+    float smallest = std::numeric_limits<float>::max();
+    std::ptrdiff_t idx = -1;
+    float boxHeight = 0.0f;
+    const ::cv::Point2f cc = center(current);
+    for (std::size_t i = 0; i < boxes.size(); ++i) {
+        if (ignored.contains(i)) {
+            continue;
+        }
+        const ::cv::Point2f pc = center(boxes[i]);
+        const float d = dist(cc, pc);
+        if (d >= smallest) {
+            continue;
+        }
+        const float h = minSide(boxes[i]);
+        const float lineDistance =
+            isVertical ? std::fabs(pc.x - (m * pc.y + c)) : std::fabs(pc.y - (m * pc.x + c));
+        if (lineDistance < h * centerThreshold) {
+            idx = static_cast<std::ptrdiff_t>(i);
+            smallest = d;
+            boxHeight = h;
+        }
+    }
+    if (idx == -1) {
+        return std::nullopt;
+    }
+    return std::make_pair(static_cast<std::size_t>(idx), boxHeight);
+}
+
+Box mergeBoxes(const Box &a, const Box &b) {
+    return {std::min(a.x0, b.x0), std::min(a.y0, b.y0), std::max(a.x1, b.x1), std::max(a.y1, b.y1),
+            a.angle};
+}
+
+// CRAFT box grouping -> reading-ordered text lines.
+std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
+                                float distanceThreshold, float heightThreshold,
+                                float minSideThreshold, float maxSideThreshold,
+                                float verticalThreshold) {
+    std::sort(boxes.begin(), boxes.end(),
+              [](const Box &a, const Box &b) { return maxSide(a) > maxSide(b); });
+
+    std::vector<Box> merged;
+    std::unordered_set<std::size_t> ignored;
+    while (!boxes.empty()) {
+        Box current = boxes.front();
+        const float normalizedAngle = (current.angle > 45.0f) ? current.angle - 90.0f : current.angle;
+        boxes.erase(boxes.begin());
+        ignored.clear();
+        float lineAngle = 0.0f;
+
+        while (true) {
+            auto [slope, intercept, isVertical] = fitLineToShortestSides(current, verticalThreshold);
+            lineAngle = isVertical ? -90.0f : std::atan(slope) * 180.0f / static_cast<float>(M_PI);
+            auto closest =
+                findClosestBox(boxes, ignored, current, isVertical, slope, intercept, centerThreshold);
+            if (!closest) {
+                break;
+            }
+            const auto [candIdx, candHeight] = *closest;
+            Box candidate = boxes[candIdx];
+            if ((isClose(candidate.angle, 90.0f) && !isVertical) ||
+                (isClose(candidate.angle, 0.0f) && isVertical)) {
+                candidate = rotateBox(candidate, normalizedAngle);
+            }
+            const float md = minDistanceBetween(candidate, current);
+            const float mergedHeight = minSide(current);
+            if (md < distanceThreshold * candHeight &&
+                std::fabs(mergedHeight - candHeight) < candHeight * heightThreshold) {
+                current = mergeBoxes(current, candidate);
+                boxes.erase(boxes.begin() + static_cast<std::ptrdiff_t>(candIdx));
+                ignored.clear();
+            } else {
+                ignored.insert(candIdx);
+            }
+        }
+        current.angle = lineAngle;
+        merged.push_back(current);
+    }
+
+    // remove small boxes
+    std::vector<Box> filtered;
+    for (const auto &b : merged) {
+        if (minSide(b) > minSideThreshold && maxSide(b) > maxSideThreshold) {
+            filtered.push_back(b);
+        }
+    }
+
+    // reading order: rows by top-Y, then left-to-right within a row
+    std::sort(filtered.begin(), filtered.end(),
+              [](const Box &a, const Box &b) { return a.y0 < b.y0; });
+    float yThresh = 0.0f;
+    if (!filtered.empty()) {
+        float total = 0.0f;
+        for (const auto &b : filtered) {
+            total += minSide(b);
+        }
+        yThresh = (total / static_cast<float>(filtered.size())) * 0.5f;
+    }
+    for (auto rowBegin = filtered.begin(); rowBegin != filtered.end();) {
+        const float rowY = rowBegin->y0;
+        auto rowEnd = std::find_if(rowBegin, filtered.end(),
+                                   [rowY, yThresh](const Box &b) { return b.y0 - rowY > yThresh; });
+        std::sort(rowBegin, rowEnd, [](const Box &a, const Box &b) { return a.x0 < b.x0; });
+        rowBegin = rowEnd;
+    }
+    return filtered;
+}
+
+// Char-level CRAFT extraction: one upright box per glyph, no line grouping. The
+// affinity map is SUBTRACTED from the text map to break the links between
+// adjacent characters (the opposite of the grouped path, which adds them), then
+// the components are eroded/dilated to clean up before labelling. Used by the
+// second, per-column detection pass that reads upright stacked text glyph by
+// glyph. Mirrors the old VerticalDetector's single-character path.
+std::vector<Box> getCharBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityMap,
+                                         float textThreshold, float linkThreshold,
+                                         float lowTextThreshold) {
+    const int32_t imgH = textMap.rows;
+    const int32_t imgW = textMap.cols;
+    ::cv::Mat textScore, affinityScore;
+    ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
+    ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
+                    ::cv::THRESH_BINARY);
+    ::cv::Mat comb = textScore - affinityScore; // subtract to separate adjacent glyphs
+    ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_TOZERO);
+    ::cv::threshold(comb, comb, 1.0, 1.0, ::cv::THRESH_TRUNC);
+    ::cv::Mat kernel = ::cv::getStructuringElement(::cv::MORPH_RECT, ::cv::Size(3, 3));
+    ::cv::erode(comb, comb, kernel, ::cv::Point(-1, -1), 1);
+    ::cv::dilate(comb, comb, kernel, ::cv::Point(-1, -1), 4);
+
+    ::cv::Mat binary;
+    comb.convertTo(binary, CV_8UC1);
+    ::cv::Mat labels, stats, centroids;
+    const int32_t nLabels = ::cv::connectedComponentsWithStats(binary, labels, stats, centroids, 4);
+
+    std::vector<Box> boxes;
+    boxes.reserve(static_cast<std::size_t>(nLabels));
+    for (int32_t i = 1; i < nLabels; ++i) {
+        auto box = boxFromComponent(textMap, labels, stats, i, imgW, imgH, lowTextThreshold);
+        if (box) {
+            box->angle = 0.0f; // glyphs are read upright, never rotated
+            boxes.push_back(*box);
+        }
+    }
+    return boxes;
+}
+
+// CRAFT half-res heatmap (text+affinity interleaved) -> oriented quads in
+// detector-input pixels; restoreRatio scales the half-res boxes back up. With
+// charLevel the boxes are individual upright glyphs (no grouping); otherwise
+// they are grouped reading-ordered lines.
+std::vector<Quad> extractCraft(std::span<const float> data, int32_t heatW, int32_t heatH,
+                               float textThreshold, float linkThreshold, float lowTextThreshold,
+                               float restoreRatio, bool charLevel) {
+    auto [textMap, affinityMap] = interleavedToMats(data, ::cv::Size(heatW, heatH));
+    std::vector<Box> boxes =
+        charLevel ? getCharBoxesFromTextMap(textMap, affinityMap, textThreshold, linkThreshold,
+                                            lowTextThreshold)
+                  : getDetBoxesFromTextMap(textMap, affinityMap, textThreshold, linkThreshold,
+                                           lowTextThreshold);
+    for (auto &b : boxes) {
+        b.x0 *= restoreRatio;
+        b.y0 *= restoreRatio;
+        b.x1 *= restoreRatio;
+        b.y1 *= restoreRatio;
+    }
+    if (!charLevel) {
+        // Grouping constants in detector-input space. Lines are merged without a width
+        // cap; the recognizer reads each line whole (snapping to its widest bucket).
+        boxes = groupTextBoxes(boxes, /*center*/ 0.5f, /*distance*/ 2.0f, /*height*/ 2.0f,
+                               /*minSide*/ 15.0f, /*maxSide*/ 30.0f,
+                               /*verticalThreshold*/ 20.0f);
+    }
+
+    std::vector<Quad> quads;
+    quads.reserve(boxes.size());
+    for (const auto &b : boxes) {
+        Quad q;
+        q.score = 1.0f;
+        q.angle = b.angle;
+        // De-skew near-horizontal lines by rotating the AABB corners about the
+        // center. A near-vertical line (angle ~ -90, from a tall/stacked region)
+        // is NOT flipped flat — that would lay an upright column on its side and
+        // misplace the box; keep it as an upright tall AABB so the column reader
+        // can take it.
+        const ::cv::Point2f ctr = center(b);
+        const float rad =
+            (std::fabs(b.angle) > 45.0f) ? 0.0f : b.angle * static_cast<float>(M_PI) / 180.0f;
+        const auto cs = corners(b);
+        for (std::size_t i = 0; i < 4; ++i) {
+            q.pts[i] = rotateAround(cs[i], ctr, rad);
+        }
+        quads.push_back(q);
+    }
+    return quads;
+}
+
+// ------------------------------ DBNet branch -------------------------------
+// DBNet prob map [H,W] -> oriented quads.
+std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, float boxThreshold,
+                               float unclipRatio, int32_t minBoxSide, int32_t maxCandidates,
+                               bool applySigmoid) {
+    const int32_t w = probIn.cols;
+    const int32_t h = probIn.rows;
+    // The caller declares (from the model's export contract) whether the head
+    // emits raw logits (apply sigmoid) or already-normalized probabilities.
+    ::cv::Mat prob;
+    if (applySigmoid) {
+        ::cv::Mat neg;
+        ::cv::exp(-probIn, neg);
+        prob = 1.0 / (1.0 + neg);
+    } else {
+        prob = probIn;
+    }
+
+    ::cv::Mat bitmap;
+    ::cv::threshold(prob, bitmap, static_cast<double>(binThreshold), 255, ::cv::THRESH_BINARY);
+    bitmap.convertTo(bitmap, CV_8UC1);
+
+    std::vector<std::vector<::cv::Point>> contours;
+    ::cv::findContours(bitmap, contours, ::cv::RETR_LIST, ::cv::CHAIN_APPROX_SIMPLE);
+
+    std::vector<Quad> quads;
+    const int32_t maxN = static_cast<int32_t>(
+        std::min<std::size_t>(contours.size(), static_cast<std::size_t>(maxCandidates)));
+    for (int32_t i = 0; i < maxN; ++i) {
+        const auto &contour = contours[static_cast<std::size_t>(i)];
+        if (contour.size() < 4) {
+            continue;
+        }
+        ::cv::RotatedRect rr = ::cv::minAreaRect(contour);
+        if (std::min(rr.size.width, rr.size.height) < static_cast<float>(minBoxSide)) {
+            continue;
+        }
+        ::cv::Mat mask = ::cv::Mat::zeros(prob.size(), CV_8UC1);
+        ::cv::drawContours(mask, contours, i, ::cv::Scalar(255), ::cv::FILLED);
+        const float score = static_cast<float>(::cv::mean(prob, mask)[0]);
+        if (score < boxThreshold) {
+            continue;
+        }
+        const double area = static_cast<double>(rr.size.width) * static_cast<double>(rr.size.height);
+        const double perim =
+            2.0 * (static_cast<double>(rr.size.width) + static_cast<double>(rr.size.height));
+        const double distance = perim > 0.0 ? area * static_cast<double>(unclipRatio) / perim : 0.0;
+        const auto grow = static_cast<float>(2.0 * distance);
+        ::cv::RotatedRect expanded(rr.center,
+                                   ::cv::Size2f(rr.size.width + grow, rr.size.height + grow),
+                                   rr.angle);
+        if (std::min(expanded.size.width, expanded.size.height) <
+            static_cast<float>(minBoxSide + 2)) {
+            continue;
+        }
+        ::cv::Point2f c[4];
+        expanded.points(c);
+        Quad q;
+        q.score = score;
+        q.angle = expanded.angle;
+        float minX = static_cast<float>(w), minY = static_cast<float>(h), maxX = 0, maxY = 0;
+        for (int32_t k = 0; k < 4; ++k) {
+            const float px = std::clamp(c[k].x, 0.0f, static_cast<float>(w));
+            const float py = std::clamp(c[k].y, 0.0f, static_cast<float>(h));
+            q.pts[static_cast<std::size_t>(k)] = {px, py};
+            minX = std::min(minX, px);
+            minY = std::min(minY, py);
+            maxX = std::max(maxX, px);
+            maxY = std::max(maxY, py);
+        }
+        if (maxX - minX < 1.0f || maxY - minY < 1.0f) {
+            continue;
+        }
+        quads.push_back(q);
+    }
+
+    std::sort(quads.begin(), quads.end(), [](const Quad &a, const Quad &b) {
+        if (std::abs(a.pts[0].y - b.pts[0].y) > 10.0f) {
+            return a.pts[0].y < b.pts[0].y;
+        }
+        return a.pts[0].x < b.pts[0].x;
+    });
+    return quads;
+}
+
+// ----------------------------- option readers ------------------------------
+// Required option getters (defaults live in the TS wrappers, so these throw).
+double getNumberProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isNumber()) {
+        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a number");
+    }
+    return opts.getProperty(rt, name).asNumber();
+}
+
+std::string getStringProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isString()) {
+        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a string");
+    }
+    return opts.getProperty(rt, name).asString(rt).utf8(rt);
+}
+
+bool getBoolProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
+        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a boolean");
+    }
+    return opts.getProperty(rt, name).asBool();
+}
+
+// Optional boolean (defaults when absent) — used for flags a caller may omit.
+bool getBoolPropOr(jsi::Runtime &rt, const jsi::Object &opts, const char *name, bool fallback) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
+        return fallback;
+    }
+    return opts.getProperty(rt, name).asBool();
+}
+
+// Flatten quads to a JS double array, 10 per box (x0,y0..x3,y3,score,angle).
+jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
+    jsi::Array out(rt, quads.size() * 10);
+    size_t idx = 0;
+    for (const auto &q : quads) {
+        for (std::size_t k = 0; k < 4; ++k) {
+            out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.pts[k].x)));
+            out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.pts[k].y)));
+        }
+        out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.score)));
+        out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.angle)));
+    }
+    return out;
+}
+
+} // namespace
+
+void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
+    auto name = "extractTextBoxes";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 2) {
+            throw jsi::JSError(rt, "Usage: extractTextBoxes(src, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "extractTextBoxes: src must be a Tensor");
+        }
+        if (!args[1].isObject()) {
+            throw jsi::JSError(rt, "extractTextBoxes: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto opts = args[1].asObject(rt);
+
+        if (src->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "extractTextBoxes: src must be a float32 Tensor");
+        }
+        if (!opts.hasProperty(rt, "mode") || !opts.getProperty(rt, "mode").isString()) {
+            throw jsi::JSError(rt, "extractTextBoxes: options.mode is required and must be a string");
+        }
+        const std::string mode = opts.getProperty(rt, "mode").asString(rt).utf8(rt);
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "extractTextBoxes: src tensor is currently in use");
+        }
+        if (!src->data_) {
+            throw jsi::JSError(rt, "extractTextBoxes: src tensor has been disposed");
+        }
+
+        const auto *dataPtr = reinterpret_cast<const float *>(src->data_.get());
+        std::span<const float> data(dataPtr, src->numel_);
+
+        std::vector<Quad> quads;
+        try {
+            if (mode == "craft") {
+                // src is [1,Hd,Wd,2] or [Hd,Wd,2] interleaved (text, affinity), half-res.
+                const auto &s = src->shape_;
+                if (s.size() < 3 || s.back() != 2) {
+                    throw jsi::JSError(rt, "extractTextBoxes: craft src must be [..,Hd,Wd,2]");
+                }
+                const int32_t heatW = s[s.size() - 2];
+                const int32_t heatH = s[s.size() - 3];
+                const double targetH = getNumberProp(rt, opts, "targetHeight");
+                const float restoreRatio = static_cast<float>(targetH) / static_cast<float>(heatH);
+                quads = extractCraft(data, heatW, heatH,
+                                     static_cast<float>(getNumberProp(rt, opts, "textThreshold")),
+                                     static_cast<float>(getNumberProp(rt, opts, "linkThreshold")),
+                                     static_cast<float>(getNumberProp(rt, opts, "lowTextThreshold")),
+                                     restoreRatio, getBoolPropOr(rt, opts, "charLevel", false));
+            } else if (mode == "dbnet") {
+                // src is [1,1,H,W] or [H,W] probability map (full-res).
+                const auto &s = src->shape_;
+                if (s.size() < 2) {
+                    throw jsi::JSError(rt, "extractTextBoxes: dbnet src must be [..,H,W]");
+                }
+                const int32_t w = s[s.size() - 1];
+                const int32_t h = s[s.size() - 2];
+                ::cv::Mat prob(h, w, CV_32F, const_cast<float *>(dataPtr));
+                quads = extractDbnet(
+                    prob, static_cast<float>(getNumberProp(rt, opts, "binThreshold")),
+                    static_cast<float>(getNumberProp(rt, opts, "boxThreshold")),
+                    static_cast<float>(getNumberProp(rt, opts, "unclipRatio")),
+                    static_cast<int32_t>(getNumberProp(rt, opts, "minBoxSide")),
+                    static_cast<int32_t>(getNumberProp(rt, opts, "maxCandidates")),
+                    getBoolProp(rt, opts, "applySigmoid"));
+            } else {
+                throw jsi::JSError(rt, "extractTextBoxes: unknown mode '" + mode + "'");
+            }
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("extractTextBoxes: OpenCV error: ") + e.what());
+        }
+        return quadsToArray(rt, quads);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             2, fnBody));
+}
+
+// ------------------------------- warpQuad ----------------------------------
+void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
+    auto name = "warpQuad";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 4) {
+            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
+        }
+        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
+        }
+        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
+            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
+        }
+        if (!args[3].isObject()) {
+            throw jsi::JSError(rt, "warpQuad: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
+        }
+        auto quadArr = args[2].asObject(rt).asArray(rt);
+        auto opts = args[3].asObject(rt);
+
+        if (quadArr.length(rt) != 8) {
+            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
+        }
+        if (src->shape_[2] != dst->shape_[2]) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
+        }
+
+        const int32_t channels = src->shape_[2];
+        const int32_t recH = dst->shape_[0];
+        const int32_t bucketW = dst->shape_[1];
+
+        if (!opts.hasProperty(rt, "contentWidth") ||
+            !opts.getProperty(rt, "contentWidth").isNumber()) {
+            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
+        }
+        const int32_t contentWidth =
+            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
+                       bucketW);
+        const std::string padMode = getStringProp(rt, opts, "padMode");
+        const double padValue = getNumberProp(rt, opts, "padValue");
+        const std::string align = getStringProp(rt, opts, "align");
+
+        std::array<::cv::Point2f, 4> quad;
+        for (std::size_t i = 0; i < 8; ++i) {
+            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
+                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
+            }
+        }
+        for (std::size_t i = 0; i < 4; ++i) {
+            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
+                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
+        }
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
+        }
+        if (!src->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
+        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
+
+        try {
+            const ::cv::Point2f dstPts[4] = {{0.0f, 0.0f},
+                                             {static_cast<float>(contentWidth), 0.0f},
+                                             {static_cast<float>(contentWidth),
+                                              static_cast<float>(recH)},
+                                             {0.0f, static_cast<float>(recH)}};
+            const ::cv::Point2f srcPts[4] = {quad[0], quad[1], quad[2], quad[3]};
+            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts, dstPts);
+            ::cv::Mat content;
+            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
+                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
+
+            ::cv::Scalar padColor;
+            if (padMode == "cornerMean") {
+                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
+                ::cv::Scalar acc(0, 0, 0, 0);
+                const std::array<::cv::Rect, 4> rects = {
+                    ::cv::Rect(0, 0, patch, patch),
+                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
+                    ::cv::Rect(0, recH - patch, patch, patch),
+                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
+                for (const auto &r : rects) {
+                    acc += ::cv::mean(content(r));
+                }
+                padColor = acc / 4.0;
+            } else {
+                padColor = ::cv::Scalar::all(padValue);
+            }
+
+            dstMat.setTo(padColor);
+            const int32_t offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
+            content.copyTo(dstMat(::cv::Rect(offsetX, 0, contentWidth, recH)));
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[1]);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             4, fnBody));
+}
+
+// --------------------------- ctcGreedyDecode -------------------------------
+void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
+    auto name = "ctcGreedyDecode";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 2) {
+            throw jsi::JSError(rt, "Usage: ctcGreedyDecode(src, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src must be a Tensor");
+        }
+        if (!args[1].isObject()) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto opts = args[1].asObject(rt);
+
+        if (src->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src must be a float32 Tensor");
+        }
+        const auto &s = src->shape_;
+        if (s.size() < 2) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src must be at least 2-D [..,T,V]");
+        }
+        const int32_t vocab = s.back();
+        if (vocab < 1) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: vocab dimension must be >= 1");
+        }
+        if (src->numel_ % static_cast<std::size_t>(vocab) != 0) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: numel must be a multiple of the vocab dim");
+        }
+        const int32_t timesteps = static_cast<int32_t>(src->numel_) / vocab;
+        const bool softmax = getBoolProp(rt, opts, "softmax");
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor is currently in use");
+        }
+        if (!src->data_) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor has been disposed");
+        }
+        const auto *data = reinterpret_cast<const float *>(src->data_.get());
+
+        jsi::Array out(rt, static_cast<size_t>(timesteps) * 2);
+        size_t oi = 0;
+        for (int32_t t = 0; t < timesteps; ++t) {
+            const float *row = data + static_cast<std::size_t>(t) * static_cast<std::size_t>(vocab);
+            const float *maxIt = std::max_element(row, row + vocab);
+            const int32_t maxIdx = static_cast<int32_t>(maxIt - row);
+            const float maxVal = *maxIt;
+            double prob = static_cast<double>(maxVal);
+            if (softmax) {
+                double sum = 0.0;
+                for (int32_t v = 0; v < vocab; ++v) {
+                    sum += std::exp(static_cast<double>(row[v]) - static_cast<double>(maxVal));
+                }
+                prob = sum > 0.0 ? 1.0 / sum : 0.0; // exp(maxVal - maxVal) / sum
+            }
+            out.setValueAtIndex(rt, oi++, jsi::Value(static_cast<double>(maxIdx)));
+            out.setValueAtIndex(rt, oi++, jsi::Value(prob));
+        }
+        return out;
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             2, fnBody));
+}
+
+} // namespace rnexecutorch::extensions::cv::ocr_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
new file mode 100644
index 0000000000..dc7905fe99
--- /dev/null
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
@@ -0,0 +1,17 @@
+#pragma once
+
+#include <jsi/jsi.h>
+
+namespace rnexecutorch::extensions::cv::ocr_ops {
+// Detector heatmap -> flat array of oriented quads (10 doubles/box:
+// x0,y0..x3,y3,score,angle), in detector-input pixels.
+void install_extractTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+
+// Perspective-crop an oriented quad of `src` into the recognizer canvas `dst`
+// (crop + resize-to-height + bucket-pad), HWC uint8.
+void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+
+// Per-timestep argmax + max value over [..,T,V] logits -> flat [idx,prob,...];
+// options.softmax makes `prob` a probability.
+void install_ctcGreedyDecode(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+} // namespace rnexecutorch::extensions::cv::ocr_ops
diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index 7eafc57abf..89e4d70548 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -1301,3 +1301,105 @@ export type BlazeFaceLandmark = (typeof BLAZEFACE_LANDMARKS)[number];
  * @category Types
  */
 export type CocoLandmark = (typeof COCO_LANDMARKS)[number];
+export { alphabets, symbols, PPOCR_SYMBOLS } from './ocrSymbols';
+export type { OCRLanguage } from './ocrSymbols';
+
+/**
+ * PP-DocLayoutV3 region classes, in model output order (index = class id). Some
+ * display names repeat at different ids — that is the model's own id2label, kept
+ * positional so `DOC_LAYOUT_LABELS[id]` matches the logits dimension.
+ * @category Constants
+ */
+export const DOC_LAYOUT_LABELS = [
+  'abstract',
+  'algorithm',
+  'aside_text',
+  'chart',
+  'content',
+  'formula',
+  'doc_title',
+  'figure_title',
+  'footer',
+  'footer',
+  'footnote',
+  'formula_number',
+  'header',
+  'header',
+  'image',
+  'formula',
+  'number',
+  'paragraph_title',
+  'reference',
+  'reference_content',
+  'seal',
+  'table',
+  'text',
+  'text',
+  'vision_footnote',
+] as const;
+
+/**
+ * Type representing a valid PP-DocLayout region class string.
+ * @category Types
+ */
+export type DocLayoutLabel = (typeof DOC_LAYOUT_LABELS)[number];
+
+/**
+ * SLANet_plus table-structure token vocabulary (50 tokens; index = token id).
+ * `0:'sos'`, `49:'eos'`. The table decoder emits the HTML `<tr>/<td>` skeleton
+ * (with col/row spans) one token at a time; concatenating the non-special tokens
+ * yields the table structure HTML.
+ * @category Constants
+ */
+export const SLANET_STRUCTURE_VOCAB = [
+  'sos',
+  '<thead>',
+  '</thead>',
+  '<tbody>',
+  '</tbody>',
+  '<tr>',
+  '</tr>',
+  '<td',
+  '>',
+  '</td>',
+  ' colspan="2"',
+  ' colspan="3"',
+  ' colspan="4"',
+  ' colspan="5"',
+  ' colspan="6"',
+  ' colspan="7"',
+  ' colspan="8"',
+  ' colspan="9"',
+  ' colspan="10"',
+  ' colspan="11"',
+  ' colspan="12"',
+  ' colspan="13"',
+  ' colspan="14"',
+  ' colspan="15"',
+  ' colspan="16"',
+  ' colspan="17"',
+  ' colspan="18"',
+  ' colspan="19"',
+  ' colspan="20"',
+  ' rowspan="2"',
+  ' rowspan="3"',
+  ' rowspan="4"',
+  ' rowspan="5"',
+  ' rowspan="6"',
+  ' rowspan="7"',
+  ' rowspan="8"',
+  ' rowspan="9"',
+  ' rowspan="10"',
+  ' rowspan="11"',
+  ' rowspan="12"',
+  ' rowspan="13"',
+  ' rowspan="14"',
+  ' rowspan="15"',
+  ' rowspan="16"',
+  ' rowspan="17"',
+  ' rowspan="18"',
+  ' rowspan="19"',
+  ' rowspan="20"',
+  '<td></td>',
+  'eos',
+] as const;
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index 8360472980..a783ce3443 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,7 +1,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
-import { scalePoint } from './points';
+import { scalePoint, type Point } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
@@ -125,6 +125,28 @@ export function scaleBox<F extends BoxFormat>(
   }
 }
 
+/**
+ * Computes the axis-aligned bounding box (`xyxy`) enclosing a set of points,
+ * e.g. the corners of an oriented OCR {@link Quad}.
+ * @category Utils
+ * @param points The points to bound (need not be ordered).
+ * @returns The enclosing bounding box in `xyxy` format.
+ */
+export function boundingBoxOf(points: readonly Point[]): BoundingBox<'xyxy'> {
+  'worklet';
+  let xmin = Infinity;
+  let ymin = Infinity;
+  let xmax = -Infinity;
+  let ymax = -Infinity;
+  for (const p of points) {
+    if (p.x < xmin) xmin = p.x;
+    if (p.y < ymin) ymin = p.y;
+    if (p.x > xmax) xmax = p.x;
+    if (p.y > ymax) ymax = p.y;
+  }
+  return { format: 'xyxy', xmin, ymin, xmax, ymax };
+}
+
 /**
  * Options for Non-Maximum Suppression (NMS).
  * @category Types
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index 1d98b73c41..ed30273f4c 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -1,6 +1,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ImageFormat } from '../image';
+import type { Point } from './points';
 
 /**
  * Supported color conversion code presets (similar to OpenCV).
@@ -209,3 +210,194 @@ export function applyColormap(
   'worklet';
   return rnexecutorchJsi.cv.applyColormap(src, dst, colormap);
 }
+
+/**
+ * Detector heatmap postprocessing mode. `'craft'` consumes a half-resolution
+ * interleaved `[..,Hd,Wd,2]` (text, affinity) map; `'dbnet'` consumes a
+ * full-resolution `[..,H,W]` probability map.
+ * @category Types
+ */
+export type TextBoxMode = 'craft' | 'dbnet';
+
+/**
+ * An oriented quadrilateral text region returned by {@link extractTextBoxes},
+ * in detector-input pixel space.
+ * @category Types
+ */
+export type Quad = {
+  /** Four corner points, ordered TL, TR, BR, BL. */
+  readonly points: readonly Point[];
+  /** Detector confidence for this region (1 for CRAFT grouped lines). */
+  readonly score: number;
+  /** Line angle in degrees. */
+  readonly angle: number;
+};
+
+/**
+ * CRAFT-mode options for {@link extractTextBoxes} (region+affinity heatmap).
+ * @category Types
+ */
+export type CraftExtractOptions = {
+  readonly mode: 'craft';
+  readonly textThreshold: number;
+  readonly linkThreshold: number;
+  readonly lowTextThreshold: number;
+  /**
+   * Detector input height — used to restore the half-res boxes. Required (the
+   * native op throws without it); the OCR pipeline sets it per run.
+   */
+  readonly targetHeight: number;
+  /**
+   * Return one upright box per glyph instead of grouped reading-ordered lines:
+   * the affinity map is subtracted (not added) to separate adjacent characters,
+   * and no line grouping or de-skew rotation is applied. Used by the second,
+   * per-column pass that reads upright stacked text. Defaults to `false`.
+   */
+  readonly charLevel?: boolean;
+};
+
+/**
+ * DBNet-mode options for {@link extractTextBoxes} (probability map).
+ * @category Types
+ */
+export type DbnetExtractOptions = {
+  readonly mode: 'dbnet';
+  readonly binThreshold: number;
+  readonly boxThreshold: number;
+  readonly unclipRatio: number;
+  readonly minBoxSide: number;
+  readonly maxCandidates: number;
+  /**
+   * Whether the probability map needs a sigmoid applied first. Set from the
+   * model's export contract (`true` = raw logits, `false` = already probabilities).
+   */
+  readonly applySigmoid: boolean;
+};
+
+/**
+ * Options for {@link extractTextBoxes} — a discriminated union on `mode`, so the
+ * type system enforces exactly the field set the chosen detector needs.
+ * @category Types
+ */
+export type ExtractTextBoxesOptions = CraftExtractOptions | DbnetExtractOptions;
+
+/**
+ * Converts a detector heatmap tensor into oriented text-line quads.
+ *
+ * Variable-sized output: the native op returns a flat array of 10 doubles per
+ * box which this wrapper reshapes into {@link Quad}s.
+ * @category Typescript API
+ * @param src The detector output tensor (float32). CRAFT: `[..,Hd,Wd,2]`;
+ * DBNet: `[..,H,W]`.
+ * @param opts Mode and per-mode thresholds.
+ * @returns The detected quads in detector-input pixel space.
+ */
+export function extractTextBoxes(src: Tensor, opts: ExtractTextBoxesOptions): Quad[] {
+  'worklet';
+  const flat = rnexecutorchJsi.cv.extractTextBoxes(src, opts) as number[];
+  const quads: Quad[] = [];
+  for (let i = 0; i < flat.length; i += 10) {
+    quads.push({
+      points: [
+        { x: flat[i]!, y: flat[i + 1]! },
+        { x: flat[i + 2]!, y: flat[i + 3]! },
+        { x: flat[i + 4]!, y: flat[i + 5]! },
+        { x: flat[i + 6]!, y: flat[i + 7]! },
+      ],
+      score: flat[i + 8]!,
+      angle: flat[i + 9]!,
+    });
+  }
+  return quads;
+}
+
+/**
+ * Options for {@link warpQuad}.
+ * @category Types
+ */
+export type WarpQuadOptions = {
+  /** Width (px) of the warped content inside the destination canvas. */
+  readonly contentWidth: number;
+  /** Horizontal placement of the content in the canvas. Default `'left'`. */
+  readonly align?: 'left' | 'center';
+  /** How the remaining canvas is filled. Default `'constant'`. */
+  readonly padMode?: 'constant' | 'cornerMean';
+  /** Fill value for `padMode: 'constant'`. Default `0`. */
+  readonly padValue?: number;
+};
+
+/**
+ * Perspective-crops an oriented quad region of `src` into the pre-allocated
+ * recognizer canvas `dst`, folding crop + resize-to-height + bucket-pad.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout. Shape [H,W,C].
+ * @param dst The pre-allocated recognizer canvas in HWC uint8 layout, sized
+ * `[recH, bucketW, C]`.
+ * @param quad Eight numbers `[x0,y0,..,x3,y3]` (TL,TR,BR,BL) in `src` pixels.
+ * @param opts Content width, alignment, and padding configuration.
+ * @returns The destination tensor `dst`.
+ */
+export function warpQuad(src: Tensor, dst: Tensor, quad: number[], opts: WarpQuadOptions): Tensor {
+  'worklet';
+  // Defaults live here (the native op requires every field) — see add-native-extension.
+  return rnexecutorchJsi.cv.warpQuad(src, dst, quad, {
+    contentWidth: opts.contentWidth,
+    align: opts.align ?? 'left',
+    padMode: opts.padMode ?? 'constant',
+    padValue: opts.padValue ?? 0,
+  });
+}
+
+/**
+ * Options for {@link ctcGreedyDecode}.
+ * @category Types
+ */
+export type CtcGreedyDecodeOptions = {
+  /**
+   * Apply a per-timestep softmax so `values` are probabilities (CRNN logits).
+   * Leave `false` for heads that are already softmaxed (SVTR).
+   */
+  readonly softmax?: boolean;
+};
+
+/**
+ * Per-timestep argmax + max value over recognizer logits, computed natively on
+ * the tensor buffer (avoids copying the whole `[T, V]` tensor into JS).
+ * @category Typescript API
+ * @param src The recognizer output tensor (float32), shape `[..,T,V]`.
+ * @param opts Whether to softmax each timestep before taking the max value.
+ * @returns The per-timestep argmax `indices` and (optionally softmaxed) max
+ * `values`.
+ */
+export function ctcGreedyDecode(
+  src: Tensor,
+  opts?: CtcGreedyDecodeOptions
+): { indices: number[]; values: number[] } {
+  'worklet';
+  const flat = rnexecutorchJsi.cv.ctcGreedyDecode(src, {
+    softmax: opts?.softmax ?? false,
+  }) as number[];
+  const indices: number[] = [];
+  const values: number[] = [];
+  for (let i = 0; i < flat.length; i += 2) {
+    indices.push(flat[i]!);
+    values.push(flat[i + 1]!);
+  }
+  return { indices, values };
+}
+
+/**
+ * Resamples `src` through a backward sampling field (the `torch.grid_sample`
+ * step of a geometric dewarp) into the pre-allocated `dst`, natively via
+ * `cv::remap`.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
+ * @param grid The sampling field tensor (float32), shape `[..,2,gH,gW]`, channel
+ * 0 = x and 1 = y, normalized to `[-1, 1]` with `align_corners=true`.
+ * @param dst The pre-allocated destination tensor, same shape/dtype as `src`.
+ * @returns The destination tensor `dst`.
+ */
+export function gridSample(src: Tensor, grid: Tensor, dst: Tensor): Tensor {
+  'worklet';
+  return rnexecutorchJsi.cv.gridSample(src, grid, dst);
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 4464061e18..7114ffe04b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -9,6 +9,19 @@ export type Point = {
   readonly y: number;
 };
 
+/**
+ * Clamps a scalar to the inclusive range `[lo, hi]`.
+ * @category Utils
+ * @param v The value to clamp.
+ * @param lo The lower bound.
+ * @param hi The upper bound.
+ * @returns `v` constrained to `[lo, hi]`.
+ */
+export function clamp(v: number, lo: number, hi: number): number {
+  'worklet';
+  return v < lo ? lo : v > hi ? hi : v;
+}
+
 /**
  * Helper function to scale a 2D point based on resize mode and resolution
  * changes.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts
new file mode 100644
index 0000000000..c35d40a434
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts
@@ -0,0 +1,179 @@
+import type { ImageBuffer } from '../image';
+import { FORMAT_CHANNELS } from '../ops/image';
+import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
+import { clamp } from '../ops/points';
+import type { OCRDetection } from './ocr';
+
+// Crops an axis-aligned region out of an ImageBuffer (pure pixel slice, same
+// format). Used to feed a layout region (e.g. a table) to another model.
+export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>): ImageBuffer {
+  'worklet';
+  const { data, width, height, format } = input;
+  const ch = FORMAT_CHANNELS[format];
+  const x0 = clamp(Math.round(bbox.xmin), 0, width);
+  const y0 = clamp(Math.round(bbox.ymin), 0, height);
+  const x1 = clamp(Math.round(bbox.xmax), 0, width);
+  const y1 = clamp(Math.round(bbox.ymax), 0, height);
+  const cw = Math.max(1, x1 - x0);
+  const chh = Math.max(1, y1 - y0);
+  const out = new Uint8Array(cw * chh * ch);
+  for (let y = 0; y < chh; y++) {
+    const srcStart = ((y0 + y) * width + x0) * ch;
+    out.set(data.subarray(srcStart, srcStart + cw * ch), y * cw * ch);
+  }
+  return { data: out, width: cw, height: chh, format, layout: input.layout };
+}
+
+// Rotates an ImageBuffer clockwise by 0/90/180/270 degrees (pure pixel reindex).
+export function rotateImageBuffer(input: ImageBuffer, degCW: 0 | 90 | 180 | 270): ImageBuffer {
+  'worklet';
+  if (degCW === 0) {
+    return input;
+  }
+  const { data, width, height, format } = input;
+  const ch = FORMAT_CHANNELS[format];
+  const swap = degCW === 90 || degCW === 270;
+  const ow = swap ? height : width;
+  const oh = swap ? width : height;
+  const out = new Uint8Array(ow * oh * ch);
+  for (let y = 0; y < height; y++) {
+    for (let x = 0; x < width; x++) {
+      let ox = 0;
+      let oy = 0;
+      if (degCW === 90) {
+        ox = height - 1 - y;
+        oy = x;
+      } else if (degCW === 180) {
+        ox = width - 1 - x;
+        oy = height - 1 - y;
+      } else {
+        // 270
+        ox = y;
+        oy = width - 1 - x;
+      }
+      const si = (y * width + x) * ch;
+      const di = (oy * ow + ox) * ch;
+      for (let c = 0; c < ch; c++) {
+        out[di + c] = data[si + c]!;
+      }
+    }
+  }
+  return { data: out, width: ow, height: oh, format, layout: input.layout };
+}
+
+/**
+ * Index of the maximum value in `arr[offset, offset+len)` (single pass, no
+ * allocation). Shared by the layout / orientation / table decoders.
+ * @category Typescript API
+ * @param arr The array to scan.
+ * @param offset The start index of the window.
+ * @param len The window length.
+ * @returns The index (relative to `offset`) of the maximum value.
+ */
+export function argmaxRange(arr: ArrayLike<number>, offset: number, len: number): number {
+  'worklet';
+  let idx = 0;
+  let best = arr[offset]!;
+  for (let i = 1; i < len; i++) {
+    const v = arr[offset + i]!;
+    if (v > best) {
+      best = v;
+      idx = i;
+    }
+  }
+  return idx;
+}
+
+// Assigns each value to one of `k` ordered groups by cutting the sorted values at
+// the (k-1) largest gaps, then returns the MEAN of each group (its center). Never
+// a fixed gap, which would merge dense rows.
+function splitIntoK(values: readonly number[], k: number): number[] {
+  'worklet';
+  const v = [...values].sort((a, b) => a - b);
+  if (v.length <= k) {
+    return v;
+  }
+  const cuts = v
+    .map((x, i) => ({ i, g: i ? x - v[i - 1]! : -1 }))
+    .slice(1)
+    .sort((a, b) => b.g - a.g)
+    .slice(0, k - 1)
+    .map((o) => o.i)
+    .sort((a, b) => a - b);
+  const centers: number[] = [];
+  let prev = 0;
+  for (const cut of [...cuts, v.length]) {
+    const grp = v.slice(prev, cut);
+    centers.push(grp.reduce((s, x) => s + x, 0) / grp.length);
+    prev = cut;
+  }
+  return centers;
+}
+
+// Index of the center in `cs` closest to `x`.
+function nearest(x: number, cs: readonly number[]): number {
+  'worklet';
+  let b = 0;
+  for (let j = 1; j < cs.length; j++) {
+    if (Math.abs(x - cs[j]!) < Math.abs(x - cs[b]!)) {
+      b = j;
+    }
+  }
+  return b;
+}
+
+/**
+ * Reconstructs a table as filled HTML from the SLANet structure skeleton and the
+ * region's OCR lines. Grid size comes from the structure (`<tr>` count, max `<td>`
+ * per row); box centers are split into that many rows/cols at the largest gaps and
+ * each box placed in its nearest cell (global column centers keep columns aligned).
+ * Falls back to a document-order fill when there is no grid.
+ *
+ * Geometric alignment only — dense bottom rows can misplace a value, since the
+ * structure export has no cell boxes.
+ * @category Typescript API
+ * @param html The SLANet structure HTML skeleton (gives the grid size).
+ * @param lines The table region's OCR lines (with page-space quads).
+ * @returns A `<table>` with each cell filled by its nearest-assigned text.
+ */
+export function fillTableCells(html: string, lines: readonly OCRDetection[]): string {
+  'worklet';
+  // Grid size straight from the structure: rows = <tr> count, cols = widest row.
+  const nRows = (html.match(/<tr>/g) ?? []).length;
+  let nCols = 0;
+  const trRe = /<tr>([\s\S]*?)<\/tr>/g;
+  let tr: RegExpExecArray | null;
+  while ((tr = trRe.exec(html)) !== null) {
+    nCols = Math.max(nCols, (tr[1]!.match(/<td/g) ?? []).length);
+  }
+  // No grid or nothing to place: plain document-order fill of the skeleton.
+  if (nRows === 0 || nCols === 0 || lines.length === 0) {
+    let i = 0;
+    return html.replace(/<td([^>]*)><\/td>/g, (_m, attrs) => {
+      const text = i < lines.length ? lines[i]!.text : '';
+      i++;
+      return `<td${attrs}>${text}</td>`;
+    });
+  }
+
+  const cx: number[] = [];
+  const cy: number[] = [];
+  for (const l of lines) {
+    const b = boundingBoxOf(l.quad);
+    cx.push((b.xmin + b.xmax) / 2);
+    cy.push((b.ymin + b.ymax) / 2);
+  }
+  const rowC = splitIntoK(cy, nRows);
+  const colC = splitIntoK(cx, nCols);
+  const grid: string[][] = Array.from({ length: rowC.length }, () =>
+    new Array<string>(colC.length).fill('')
+  );
+  for (let i = 0; i < lines.length; i++) {
+    const r = nearest(cy[i]!, rowC);
+    const c = nearest(cx[i]!, colC);
+    grid[r]![c] = `${grid[r]![c]!} ${lines[i]!.text}`.trim();
+  }
+  return `<table>${grid
+    .map((row) => `<tr>${row.map((t) => `<td>${t}</td>`).join('')}</tr>`)
+    .join('')}</table>`;
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
new file mode 100644
index 0000000000..ad8e751d0d
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -0,0 +1,214 @@
+import type { WorkletRuntime } from 'react-native-worklets';
+
+import { wrapAsync } from '../../../core/runtime';
+import type { ImageBuffer } from '../image';
+import type { Point } from '../ops/points';
+import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
+import { createOCR, type OCRModel, type OCRDetection } from './ocr';
+import {
+  createObjectDetector,
+  type ObjectDetectorModel,
+  type ObjectDetection,
+} from './objectDetection';
+import { createSupporting, type SupportingModel } from './supporting';
+import { cropImageBuffer, rotateImageBuffer, fillTableCells } from './documentHelpers';
+
+/**
+ * One assembled document block: a layout region (or an ungrouped catch-all) with
+ * its OCR lines grouped and concatenated, in reading order.
+ * @category Types
+ */
+export type DocumentBlock<L> = {
+  /** Region class from layout (e.g. `'text'`, `'table'`), or `'ungrouped'`. */
+  readonly regionType: L | 'ungrouped';
+  /** Block box in processing-frame pixels. */
+  readonly bbox: BoundingBox<'xyxy'>;
+  /** Layout confidence for the region (1 for `'ungrouped'`). */
+  readonly score: number;
+  /** The block's text, lines joined top-to-bottom by newlines. */
+  readonly text: string;
+  /** The OCR lines inside this block, top-to-bottom. */
+  readonly lines: readonly OCRDetection[];
+  /** Whether this block is a table region. */
+  readonly isTable: boolean;
+  /** For table blocks: the recognized HTML structure with OCR text filled in. */
+  readonly tableHtml?: string;
+};
+
+/**
+ * The result of a document OCR run.
+ * @category Types
+ */
+export type DocumentResult<L> = {
+  readonly blocks: DocumentBlock<L>[];
+  readonly regions: ObjectDetection<'xyxy', L>[];
+  readonly detections: OCRDetection[];
+  /**
+   * The frame all `bbox`/`quad` coordinates are relative to. Equals the input
+   * image unless orientation correction or dewarp was applied, in which case it
+   * is the corrected image — overlay boxes on THIS, not the original input.
+   */
+  readonly image: ImageBuffer;
+};
+
+/**
+ * Configuration for the document OCR orchestrator. Provides an OCR model, an
+ * optional layout model (regions/blocks), and an optional supporting model
+ * (orientation/dewarp pre-processing + table-structure recognition). The
+ * `orientation`/`dewarp` flags gate those pre-processing passes.
+ * @category Types
+ */
+export type DocumentOCRModel<L> = {
+  readonly ocr: OCRModel;
+  readonly layout?: ObjectDetectorModel<'xyxy', L>;
+  readonly supporting?: SupportingModel;
+  /** Detect + correct page orientation before OCR (needs `supporting`). */
+  readonly orientation?: boolean;
+  /** Geometrically dewarp the page before OCR (needs `supporting`). */
+  readonly dewarp?: boolean;
+};
+
+const centerOf = (quad: readonly Point[]): { x: number; y: number } => {
+  'worklet';
+  const b = boundingBoxOf(quad);
+  return { x: (b.xmin + b.xmax) / 2, y: (b.ymin + b.ymax) / 2 };
+};
+
+const isTableLabel = (label: unknown): boolean => {
+  'worklet';
+  return String(label) === 'table';
+};
+
+// Layout classes that carry no text — skip OCR on them.
+const VISUAL_LABELS = ['image', 'chart', 'seal'];
+const isTextRegion = (label: unknown): boolean => {
+  'worklet';
+  return !VISUAL_LABELS.includes(String(label));
+};
+
+// Shifts a crop-space detection back into page coordinates.
+function offsetDetection(d: OCRDetection, dx: number, dy: number): OCRDetection {
+  'worklet';
+  return { ...d, quad: d.quad.map((p) => ({ x: p.x + dx, y: p.y + dy })) };
+}
+
+function makeBlock<L>(
+  regionType: L | 'ungrouped',
+  bbox: BoundingBox<'xyxy'>,
+  score: number,
+  lines: OCRDetection[],
+  isTable: boolean
+): DocumentBlock<L> {
+  'worklet';
+  // Decorate with each line's y-center once, then sort top-to-bottom (avoids
+  // recomputing the bbox inside the O(n log n) comparator).
+  const sorted = lines
+    .map((line) => ({ line, cy: centerOf(line.quad).y }))
+    .sort((a, b) => a.cy - b.cy)
+    .map((e) => e.line);
+  return {
+    regionType,
+    bbox,
+    score,
+    isTable,
+    lines: sorted,
+    text: sorted.map((l) => l.text).join('\n'),
+  };
+}
+
+/**
+ * Creates the document OCR orchestrator. Pipeline: correct orientation → dewarp
+ * (supporting) → layout → per-region OCR (each text region is cropped and OCR'd on
+ * its own, upscaled into the detector — far better recall on dense pages than one
+ * whole-page pass; lines are offset back to page coords) → tables recognize their
+ * structure and fill cells with that region's OCR. Visual regions are skipped.
+ * Without layout it OCRs the whole page into one block. Layout/supporting optional.
+ * @category Typescript API
+ * @param config OCR model + optional layout + optional supporting + flags.
+ * @param runtime Optional worklet runtime thread.
+ * @returns A promise resolving to run + disposal controls.
+ */
+export async function createDocumentOCR<L>(
+  config: DocumentOCRModel<L>,
+  runtime?: WorkletRuntime
+): Promise<{
+  dispose: () => void;
+  runDocumentOCR: (input: ImageBuffer) => Promise<DocumentResult<L>>;
+  runDocumentOCRWorklet: (input: ImageBuffer) => DocumentResult<L>;
+}> {
+  const ocr = await createOCR(config.ocr, runtime);
+  const layout = config.layout
+    ? await createObjectDetector<'xyxy', L>(config.layout, runtime)
+    : null;
+  const supporting = config.supporting ? await createSupporting(config.supporting, runtime) : null;
+  const useOrientation = !!supporting && !!config.orientation;
+  const useDewarp = !!supporting && !!config.dewarp;
+
+  const dispose = () => {
+    ocr.dispose();
+    layout?.dispose();
+    supporting?.dispose();
+  };
+
+  const runDocumentOCRWorklet = (input: ImageBuffer): DocumentResult<L> => {
+    'worklet';
+    let img = input;
+    if (useOrientation && supporting) {
+      const rot = supporting.detectOrientationWorklet(img).rotationCW;
+      if (rot !== 0) {
+        img = rotateImageBuffer(img, ((360 - rot) % 360) as 0 | 90 | 180 | 270);
+      }
+    }
+    if (useDewarp && supporting) {
+      img = supporting.dewarpWorklet(img);
+    }
+
+    // Mode A — no layout: OCR the whole page into one block.
+    if (!layout) {
+      const detections = ocr.runOCRWorklet(img).detections;
+      const blocks = detections.length
+        ? [
+            makeBlock<L>(
+              'ungrouped',
+              boundingBoxOf(detections.flatMap((d) => d.quad as Point[])),
+              1,
+              detections,
+              false
+            ),
+          ]
+        : [];
+      return { blocks, regions: [], detections, image: img };
+    }
+
+    // Mode B — layout: OCR each text region's crop on its own (upscaled into the
+    // detector → far better recall than one whole-page pass), offsetting lines
+    // back to page coords. Tables also recognize structure + fill cells.
+    const regions = layout.detectObjectsWorklet(img);
+    const blocks: DocumentBlock<L>[] = [];
+    const detections: OCRDetection[] = [];
+    for (const region of regions) {
+      if (!isTextRegion(region.label)) {
+        continue;
+      }
+      const { xmin, ymin } = region.box;
+      const crop = cropImageBuffer(img, region.box);
+      const lines = ocr.runOCRWorklet(crop).detections.map((d) => offsetDetection(d, xmin, ymin));
+      const table = isTableLabel(region.label);
+      if (lines.length === 0 && !table) {
+        continue;
+      }
+      detections.push(...lines);
+      let block = makeBlock<L>(region.label, region.box, region.confidence, lines, table);
+      if (table && supporting) {
+        const structure = supporting.recognizeTableWorklet(crop);
+        block = { ...block, tableHtml: fillTableCells(structure.html, block.lines) };
+      }
+      blocks.push(block);
+    }
+    blocks.sort((a, b) => a.bbox.ymin - b.bbox.ymin || a.bbox.xmin - b.bbox.xmin);
+    return { blocks, regions, detections, image: img };
+  };
+
+  const runDocumentOCR = wrapAsync(runDocumentOCRWorklet, runtime);
+  return { runDocumentOCR, runDocumentOCRWorklet, dispose };
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
new file mode 100644
index 0000000000..70a8f3ad6c
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -0,0 +1,816 @@
+import type { WorkletRuntime } from 'react-native-worklets';
+
+import { tensor, type Tensor } from '../../../core/tensor';
+import { loadModel, type Model } from '../../../core/model';
+import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
+import { wrapAsync } from '../../../core/runtime';
+
+import type { ImageBuffer, ImageFormat } from '../image';
+import type { Point } from '../ops/points';
+import {
+  FORMAT_CONVERSION,
+  FORMAT_CHANNELS,
+  resize,
+  cvtColor,
+  toChannelsFirst,
+  normalize,
+  extractTextBoxes,
+  warpQuad,
+  ctcGreedyDecode,
+  type ColorConversionCode,
+  type CraftExtractOptions,
+  type DbnetExtractOptions,
+} from '../ops/image';
+import {
+  buildCharset,
+  mapQuadToImage,
+  orderQuad,
+  quadSize,
+  contentWidthFor,
+  flattenQuad,
+  decodeGreedy,
+  ctcConfidence,
+  nowMs,
+  snapDetectBucket,
+  snapRecognizeBucket,
+  type Buckets,
+} from './ocrHelpers';
+
+export type { Buckets } from './ocrHelpers';
+
+/**
+ * Configuration for the unified OCR pipeline. The detector and recognizer share
+ * one baked contract; only the box decoder (selected by `detectorKind`) and the
+ * default drop score differ per architecture. A model declares its architecture,
+ * its input-size buckets, and its charset.
+ * @category Types
+ */
+export type OCROptions = {
+  /**
+   * Detector architecture — selects the box decoder (CRAFT heatmap grouping vs
+   * DBNet prob-map) and the default drop score. A new architecture adds a variant.
+   */
+  readonly detectorKind: 'craft' | 'dbnet';
+  /**
+   * The model's static input-size buckets. The pipeline snaps each image to the
+   * closest `detect`/`recognize` bucket and calls the matching per-size method
+   * (`detect_<S>` / `recognize_<W>`). See {@link Buckets}.
+   */
+  readonly buckets: Buckets;
+  /**
+   * Recognizer charset (a string = one codepoint per index; an array is taken
+   * verbatim, for multi-codepoint entries like ligatures).
+   */
+  readonly charset: string | readonly string[];
+  /** Drop detections below this confidence. Defaults per detector architecture. */
+  readonly dropScore?: number;
+};
+
+/**
+ * Per-run OCR options (passed to `runOCR`, not baked into the model — toggling
+ * them needs no reload).
+ * @category Types
+ */
+export type RunOCROptions = {
+  /**
+   * Add handling for upright stacked columns (e.g. vertical signage, shipping-
+   * container codes — letters stacked top-to-bottom) on top of the normal
+   * horizontal read. X-aligned stacked glyph boxes are joined into one column
+   * word; a single tall box is cropped and its glyphs re-detected. Horizontal
+   * lines still read normally, so this only ADDS capability (at extra compute).
+   */
+  readonly vertical?: boolean;
+  /** Height/width ratio above which a box is treated as a stacked column. Default 1.5. */
+  readonly tallCropRatio?: number;
+  /** Max stacked-column re-detection passes per page (each is detector-scale). Default 8. */
+  readonly maxRedetections?: number;
+};
+
+// The unified baked contract leaves only two things per detector architecture:
+// the box decoder (selected by `detectorKind`) and the default drop score.
+// Everything else is shared — detector input is raw RGB /255 (mean/std baked in),
+// the recognizer is RGB with constant-128 left padding, both heads emit softmaxed
+// probabilities, and confidence is the mean of per-character max-probs.
+const DEFAULT_DROP_SCORE: Record<'craft' | 'dbnet', number> = { craft: 0, dbnet: 0.5 };
+
+/**
+ * Model configuration required to instantiate an OCR task runner. One fused PTE
+ * exposing `detect` + `recognize`.
+ * @category Types
+ */
+export type OCRModel = {
+  readonly modelPath: string;
+  readonly ocrOpts: OCROptions;
+};
+
+/**
+ * A single recognized text region.
+ * @category Types
+ */
+export type OCRDetection = {
+  readonly text: string;
+  readonly confidence: number;
+  /**
+   * The oriented quad (TL,TR,BR,BL) in original image pixels. Derive the
+   * axis-aligned box with `boundingBoxOf(quad)` from `cv.ops.boxes` if needed.
+   */
+  readonly quad: readonly Point[];
+  /** Wall-clock time spent recognizing this box (ms), incl. any retries. */
+  readonly recognizeMs: number;
+};
+
+/**
+ * The result of one OCR run: the recognized text regions.
+ * @category Types
+ */
+export type OCRResult = {
+  readonly detections: OCRDetection[];
+};
+
+const RECOGNIZER_ALPHA = 1 / 127.5; // (x/255 - 0.5)/0.5 -> [-1, 1]
+const RECOGNIZER_BETA = -1;
+const RECOGNIZER_PAD_VALUE = 128; // neutral gray; constant pad for both recognizers
+// Detector input is raw RGB scaled to [0,1]; the per-architecture mean/std is
+// baked into the PTE, so the client only divides by 255.
+const DETECTOR_ALPHA = 1 / 255;
+const DETECTOR_BETA = 0;
+
+// Detector box-extraction tuning. These are stable across models and not worth
+// exposing as options — bake them in. (CRAFT = EasyOCR, DBNet = PaddleOCR.)
+const CRAFT_TEXT_THRESHOLD = 0.4;
+const CRAFT_LINK_THRESHOLD = 0.4;
+const CRAFT_LOW_TEXT_THRESHOLD = 0.7;
+const DBNET_BIN_THRESHOLD = 0.3;
+const DBNET_BOX_THRESHOLD = 0.6;
+const DBNET_UNCLIP_RATIO = 1.5;
+const DBNET_MIN_BOX_SIDE = 3;
+const DBNET_MAX_CANDIDATES = 1000;
+// DBNet exports a post-sigmoid prob map, so don't re-apply sigmoid (true only for
+// raw-logit heads).
+const APPLY_SIGMOID = false;
+// A box taller than this ratio is read as an upright stacked column.
+const TALL_CROP_RATIO = 1.5;
+// Per-page cap on stacked-column re-detection passes (each is detector-scale).
+const MAX_VERTICAL_REDETECTIONS = 8;
+// Wider than this × its height = a horizontal line, never a vertical-column glyph.
+const VERTICAL_GLYPH_ASPECT = 1.6;
+// A box joins a column when its x-span overlaps the column's by this fraction and
+// the y-gap is within this × its height (loose — signage letter spacing varies).
+const VERTICAL_X_OVERLAP = 0.25;
+const VERTICAL_Y_GAP = 2.5;
+// Vertical reads are lower-confidence and opt-in, so they skip the drop-score gate.
+const VERTICAL_DROP_SCORE = 0;
+
+// cvtColor code to bring a source image format to RGB (both the detector and the
+// recognizer operate on RGB), or null when it already is RGB.
+function toRgbCode(format: ImageFormat): ColorConversionCode | null {
+  'worklet';
+  return FORMAT_CONVERSION[format].rgb;
+}
+
+// Stored at construction; CRAFT's `targetHeight` is per-run, so detectQuads adds it.
+type DetectorExtractConfig = Omit<CraftExtractOptions, 'targetHeight'> | DbnetExtractOptions;
+
+// Everything the detector pass needs, bundled so it can run both on the full
+// page and (for vertical text) on a single box crop to find its characters.
+type DetectContext = {
+  readonly model: Model;
+  readonly detectorKind: 'craft' | 'dbnet';
+  readonly detBuckets: readonly number[];
+  readonly format: ImageFormat;
+  readonly numChannels: number;
+  readonly detCode: ColorConversionCode | null;
+  readonly extractOpts: DetectorExtractConfig;
+  readonly detExtraChannels: readonly number[];
+};
+
+// Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
+// returns quads in `src` pixel space: letterbox into the snapped square bucket,
+// run `detect_<S>`, extract quads. Scratch is per-size, allocated and freed here.
+function detectQuads(
+  ctx: DetectContext,
+  src: Tensor,
+  width: number,
+  height: number,
+  charLevel = false
+): Point[][] {
+  'worklet';
+  const detS = snapDetectBucket(width, height, ctx.detBuckets);
+  // CRAFT heatmap is half-resolution; detector buckets are validated even.
+  const heat = detS / 2;
+
+  const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
+  const tDetColor = tensor('uint8', [detS, detS, 3]);
+  const tDetCF = tensor('uint8', [3, detS, detS]);
+  const tDetNorm = tensor('float32', [3, detS, detS]);
+  const tDetInput = tensor('float32', [1, 3, detS, detS]);
+  // CRAFT: half-res [1,S/2,S/2,2] heatmap; DBNet: full-res [1,1,S,S] prob map.
+  const tHeatmap =
+    ctx.detectorKind === 'dbnet'
+      ? tensor('float32', [1, 1, detS, detS])
+      : tensor('float32', [1, heat, heat, 2]);
+  const tDetExtras = ctx.detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]));
+  try {
+    src
+      .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
+      .throughIf(ctx.detCode !== null, cvtColor, tDetColor, ctx.detCode!)
+      .through(toChannelsFirst, tDetCF)
+      .through(normalize, tDetNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
+      .copyTo(tDetInput);
+
+    ctx.model.execute(`detect_${detS}`, [tDetInput], [tHeatmap, ...tDetExtras]);
+    // CRAFT needs the per-run input height to restore its half-res boxes;
+    // `charLevel` switches it to per-glyph (ungrouped) boxes for a column pass.
+    const extractOpts =
+      ctx.extractOpts.mode === 'craft'
+        ? { ...ctx.extractOpts, targetHeight: detS, charLevel }
+        : ctx.extractOpts;
+    const quads = extractTextBoxes(tHeatmap, extractOpts);
+    return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
+  } finally {
+    tDetResize.dispose();
+    tDetColor.dispose();
+    tDetCF.dispose();
+    tDetNorm.dispose();
+    tDetInput.dispose();
+    tHeatmap.dispose();
+    tDetExtras.forEach((t) => t.dispose());
+  }
+}
+
+// A recognizer width bucket's pre-allocated tensor-set (one per width in the
+// model's `recognize` buckets).
+type RecSet = {
+  readonly width: number;
+  readonly tCanvas: Tensor;
+  readonly tCF: Tensor;
+  readonly tNorm: Tensor;
+  readonly tInput: Tensor;
+  readonly tLogits: Tensor;
+};
+
+// Recognizer state for reading one quad. The source image is passed per call,
+// so the same context reads both the full page and a per-box crop.
+type RecContext = {
+  readonly model: Model;
+  readonly recSetByWidth: ReadonlyMap<number, RecSet>;
+  readonly recBuckets: readonly number[];
+  readonly recH: number;
+  readonly charset: string[];
+};
+
+// Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap content width to a
+// recognizer bucket, warp -> normalize -> execute -> greedy-CTC decode.
+function recognizeQuad(
+  ctx: RecContext,
+  src: Tensor,
+  corners: readonly Point[]
+): { text: string; conf: number } {
+  'worklet';
+  const cs = quadSize(corners);
+  const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
+  const desiredW = contentWidthFor(cs.width, cs.height, ctx.recH, maxRec);
+  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
+  // snapRecognizeBucket always returns one of recBuckets, so the set exists.
+  const rs = ctx.recSetByWidth.get(bucketW)!;
+  const contentWidth = Math.min(desiredW, bucketW);
+  warpQuad(src, rs.tCanvas, flattenQuad(corners), {
+    contentWidth,
+    align: 'left',
+    padMode: 'constant',
+    padValue: RECOGNIZER_PAD_VALUE,
+  });
+  rs.tCanvas
+    .through(toChannelsFirst, rs.tCF)
+    .through(normalize, rs.tNorm, { alpha: RECOGNIZER_ALPHA, beta: RECOGNIZER_BETA })
+    .copyTo(rs.tInput);
+  ctx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+  // Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
+  const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
+  const text = decodeGreedy(indices, ctx.charset);
+  const conf = ctcConfidence(values, indices);
+  return { text, conf };
+}
+
+// State the vertical-text path needs on top of `RecContext`: the detector (for
+// the second, character-level pass) and the page tensors it crops the box from.
+type VerticalContext = {
+  readonly detCtx: DetectContext;
+  readonly rawPage: Tensor; // full page, native format — cropped per box for re-detection
+  readonly recCode: ColorConversionCode | null; // native format -> recognizer color
+  readonly recC: number;
+  readonly tallCropRatio: number; // height/width ratio above which a box is a column
+  // Per-page budget for the (expensive) stacked-column re-detection pass.
+  readonly redetectBudget: { remaining: number };
+};
+
+// Joins glyph quads (in `src` pixel space, in reading order) into one recognizer
+// strip — each glyph warped upright to the recognizer height and laid side by
+// side — and recognizes it as a single line (joint hconcat). Returns null when
+// nothing usable was assembled.
+//
+// Must be defined BEFORE its callers: the worklet plugin captures referenced
+// worklets in source order, so a forward reference is undefined at run time.
+function recognizeGlyphStrip(
+  recCtx: RecContext,
+  src: Tensor,
+  recC: number,
+  glyphs: readonly (readonly Point[])[]
+): { text: string; conf: number } | null {
+  'worklet';
+  const recH = recCtx.recH;
+  const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+  // Warp each glyph upright to recognizer height (aspect preserved).
+  const slices: { tGlyph: Tensor; w: number }[] = [];
+  let totalW = 0;
+  for (const g of glyphs) {
+    const gs = quadSize(g);
+    if (gs.width < 1 || gs.height < 1) {
+      continue;
+    }
+    const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
+    const tGlyph = tensor('uint8', [recH, gw, recC]);
+    warpQuad(src, tGlyph, flattenQuad(g), {
+      contentWidth: gw,
+      align: 'left',
+      padMode: 'constant',
+      padValue: RECOGNIZER_PAD_VALUE,
+    });
+    slices.push({ tGlyph, w: gw });
+    totalW += gw;
+  }
+  if (slices.length === 0) {
+    return null;
+  }
+  try {
+    // Smallest bucket that fits the strip (snap up, no glyph truncated); widest
+    // bucket for very long columns.
+    const bucketW =
+      recCtx.recBuckets.find((w) => w >= totalW) ??
+      recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+    const rs = recCtx.recSetByWidth.get(bucketW)!;
+    // Assemble the strip row-major into the bucket canvas, neutral-padded.
+    const strip = new Uint8Array(recH * bucketW * recC);
+    strip.fill(RECOGNIZER_PAD_VALUE);
+    let xOff = 0;
+    for (const s of slices) {
+      if (xOff >= bucketW) {
+        break;
+      }
+      const copyW = Math.min(s.w, bucketW - xOff);
+      const glyphBytes = new Uint8Array(recH * s.w * recC);
+      s.tGlyph.getData(glyphBytes);
+      for (let oy = 0; oy < recH; oy++) {
+        const srcStart = oy * s.w * recC;
+        const row = glyphBytes.subarray(srcStart, srcStart + copyW * recC);
+        strip.set(row, (oy * bucketW + xOff) * recC);
+      }
+      xOff += s.w;
+    }
+    rs.tCanvas.setData(strip);
+    rs.tCanvas
+      .through(toChannelsFirst, rs.tCF)
+      .through(normalize, rs.tNorm, { alpha: RECOGNIZER_ALPHA, beta: RECOGNIZER_BETA })
+      .copyTo(rs.tInput);
+    recCtx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+    const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
+    const text = decodeGreedy(indices, recCtx.charset);
+    const conf = ctcConfidence(values, indices);
+    return text.length > 0 ? { text, conf } : null;
+  } finally {
+    slices.forEach((s) => s.tGlyph.dispose());
+  }
+}
+
+// Reads a single tall box that packs several stacked glyphs the detector grouped
+// into one box (e.g. a tightly-set container code): crop it upright, re-detect
+// the glyphs (char-level pass), and read them top -> bottom as a joined strip.
+// Returns null — caller falls back to a horizontal read — when the box is tiny,
+// the per-page re-detect budget is spent, or nothing is found.
+function readStackedColumn(
+  recCtx: RecContext,
+  vctx: VerticalContext,
+  ordered: readonly Point[],
+  size: { width: number; height: number }
+): { text: string; conf: number } | null {
+  'worklet';
+  const boxW = Math.round(size.width);
+  const boxH = Math.round(size.height);
+  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
+    return null;
+  }
+  vctx.redetectBudget.remaining--;
+  const recC = vctx.recC;
+  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
+  const tRecBox = tensor('uint8', [boxH, boxW, recC]);
+  try {
+    // Axis-aligned upright crop of the box, full width (no padding).
+    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
+      contentWidth: boxW,
+      align: 'left',
+      padMode: 'constant',
+      padValue: 0,
+    });
+    // Char-level second pass: per-glyph boxes (no grouping), in crop pixel space.
+    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
+    if (charQuads.length === 0) {
+      return null;
+    }
+    const boxSrc = vctx.recCode !== null ? cvtColor(tBoxRaw, tRecBox, vctx.recCode) : tBoxRaw;
+    // Stack reading order: top -> bottom by each glyph's upper edge.
+    const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
+    return recognizeGlyphStrip(recCtx, boxSrc, recC, glyphs);
+  } finally {
+    tBoxRaw.dispose();
+    tRecBox.dispose();
+  }
+}
+
+// Bounding axis-aligned quad (TL,TR,BR,BL) enclosing a set of quads.
+function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
+  'worklet';
+  let xmin = Infinity;
+  let ymin = Infinity;
+  let xmax = -Infinity;
+  let ymax = -Infinity;
+  for (const q of quads) {
+    for (const p of q) {
+      xmin = Math.min(xmin, p.x);
+      ymin = Math.min(ymin, p.y);
+      xmax = Math.max(xmax, p.x);
+      ymax = Math.max(ymax, p.y);
+    }
+  }
+  return [
+    { x: xmin, y: ymin },
+    { x: xmax, y: ymin },
+    { x: xmax, y: ymax },
+    { x: xmin, y: ymax },
+  ];
+}
+
+// Clusters glyph-like, x-aligned, stacked boxes into vertical columns; wide lines
+// and isolated boxes come back as `singles` to read normally. So `vertical` ADDS
+// column reading without disturbing horizontal reads.
+function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
+  columns: Point[][][];
+  singles: Point[][];
+} {
+  'worklet';
+  type B = {
+    q: Point[];
+    xmin: number;
+    xmax: number;
+    ymin: number;
+    ymax: number;
+    w: number;
+    h: number;
+  };
+  const candidates: B[] = [];
+  const singles: Point[][] = [];
+  for (const q of quads) {
+    let xmin = Infinity;
+    let ymin = Infinity;
+    let xmax = -Infinity;
+    let ymax = -Infinity;
+    for (const p of q) {
+      xmin = Math.min(xmin, p.x);
+      ymin = Math.min(ymin, p.y);
+      xmax = Math.max(xmax, p.x);
+      ymax = Math.max(ymax, p.y);
+    }
+    const w = xmax - xmin;
+    const h = ymax - ymin;
+    if (w > h * VERTICAL_GLYPH_ASPECT) {
+      singles.push(q as Point[]); // a horizontal line — read normally
+    } else {
+      candidates.push({ q: q as Point[], xmin, xmax, ymin, ymax, w, h });
+    }
+  }
+  // Top -> bottom, growing each column from its current bottom box. Alignment is
+  // checked against the column's accumulated x-range (not just the last box), so a
+  // narrow glyph like `I` between wider ones doesn't break the run.
+  candidates.sort((a, b) => a.ymin - b.ymin);
+  type Col = { boxes: B[]; xmin: number; xmax: number; bottom: number };
+  const cols: Col[] = [];
+  for (const b of candidates) {
+    let placed = false;
+    for (const col of cols) {
+      const overlap = Math.min(b.xmax, col.xmax) - Math.max(b.xmin, col.xmin);
+      const aligned = overlap > VERTICAL_X_OVERLAP * Math.min(b.w, col.xmax - col.xmin);
+      const gap = b.ymin - col.bottom;
+      if (aligned && gap < VERTICAL_Y_GAP * b.h && gap > -0.5 * b.h) {
+        col.boxes.push(b);
+        col.xmin = Math.min(col.xmin, b.xmin);
+        col.xmax = Math.max(col.xmax, b.xmax);
+        col.bottom = b.ymax;
+        placed = true;
+        break;
+      }
+    }
+    if (!placed) {
+      cols.push({ boxes: [b], xmin: b.xmin, xmax: b.xmax, bottom: b.ymax });
+    }
+  }
+  const columns: Point[][][] = [];
+  for (const col of cols) {
+    if (col.boxes.length >= 2) {
+      columns.push(col.boxes.map((b) => b.q)); // already top -> bottom
+    } else {
+      singles.push(col.boxes[0]!.q);
+    }
+  }
+  return { columns, singles };
+}
+
+// Reads one box that may be an upright stacked column (e.g. a shipping-container
+// code, letters stacked top-to-bottom). A normal-aspect box is read horizontally;
+// a tall box is read as a stacked column, falling back to the horizontal read.
+// `stacked` reports whether the column path produced the read (caller applies the
+// lenient vertical drop-score to those).
+function readBoxVertical(
+  recCtx: RecContext,
+  vctx: VerticalContext,
+  pageSrc: Tensor,
+  ordered: readonly Point[],
+  size: { width: number; height: number }
+): { text: string; conf: number; stacked: boolean } {
+  'worklet';
+  if (size.height >= size.width * vctx.tallCropRatio) {
+    const stacked = readStackedColumn(recCtx, vctx, ordered, size);
+    if (stacked) {
+      return { ...stacked, stacked: true };
+    }
+  }
+  return { ...recognizeQuad(recCtx, pageSrc, ordered), stacked: false };
+}
+
+// Pre-allocates one recognizer tensor-set per width bucket (each `recognize_<W>`
+// validated once) and derives the constant channel/height/vocab contract from the
+// first bucket. Kept out of the task factory; runs at construction.
+function buildRecognizerSets(
+  model: Model,
+  recBuckets: readonly number[]
+): { sets: RecSet[]; recC: number; recH: number; vocabSize: number } {
+  let recC = 0;
+  let recH = 0;
+  let vocabSize = 0;
+  const sets = recBuckets.map((w, i) => {
+    const m = validateModelSchema(
+      model,
+      `recognize_${w}`,
+      [SymbolicTensor('float32', [1, 'C', 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 'T', 'V'])]
+    );
+    const inShape = m.inputTensorMeta[0]!.shape;
+    if (i === 0) {
+      // Channels/height/vocab are constant across the width buckets.
+      recC = inShape[1]!;
+      recH = inShape[2]!;
+      vocabSize = m.outputTensorMeta[0]!.shape[2]!;
+    }
+    return {
+      width: w,
+      tCanvas: tensor('uint8', [recH, w, recC]),
+      tCF: tensor('uint8', [recC, recH, w]),
+      tNorm: tensor('float32', [recC, recH, w]),
+      tInput: tensor('float32', inShape),
+      tLogits: tensor('float32', m.outputTensorMeta[0]!.shape),
+    };
+  });
+  return { sets, recC, recH, vocabSize };
+}
+
+/**
+ * Creates a unified OCR runner for two-stage detect -> recognize models
+ * (EasyOCR / PaddleOCR). It loads one fused PTE, validates the `detect` and
+ * `recognize` methods, pre-allocates static scratch tensors sized from the
+ * model's compiled shapes, and returns recognition + disposal controls.
+ * @category Typescript API
+ * @param config OCR task configuration containing the model path and flat
+ * options.
+ * @param runtime Optional worklet runtime thread on which to run the pipeline.
+ * @returns A promise resolving to an object with recognition and disposal
+ * controls.
+ */
+export async function createOCR(
+  config: OCRModel,
+  runtime?: WorkletRuntime
+): Promise<{
+  dispose: () => void;
+  runOCR: (input: ImageBuffer, options?: RunOCROptions) => Promise<OCRResult>;
+  runOCRWorklet: (input: ImageBuffer, options?: RunOCROptions) => OCRResult;
+}> {
+  const { modelPath, ocrOpts } = config;
+  const model = await wrapAsync(loadModel, runtime)(modelPath);
+
+  const dropScore = ocrOpts.dropScore ?? DEFAULT_DROP_SCORE[ocrOpts.detectorKind];
+
+  const detBuckets = ocrOpts.buckets.detect;
+  const recBuckets = ocrOpts.buckets.recognize;
+  // Validation + scratch allocation can throw (bad buckets, missing methods,
+  // shape/charset mismatch); on any failure dispose the model and any tensors
+  // already built, so a bad config doesn't leak native memory.
+  let recSets: RecSet[] = [];
+  let recC = 3;
+  let recH = 0;
+  let charset: string[] = [];
+  let recSetByWidth: ReadonlyMap<number, RecSet> = new Map();
+  let detExtraChannels: number[] = [];
+  try {
+    if (detBuckets.length === 0 || recBuckets.length === 0) {
+      throw new Error(
+        'OCR: buckets.detect and buckets.recognize must each list at least one size.'
+      );
+    }
+    // Detector buckets feed a half-resolution CRAFT heatmap, so every side must be even.
+    if (detBuckets.some((s) => s % 2 !== 0)) {
+      throw new Error('OCR: every detect bucket side must be even (half-resolution heatmap).');
+    }
+    // Validate every detect bucket (heatmap layout is constant across sizes); keep
+    // the largest bucket's meta for the constant extra-output channels.
+    const detInSpec = [SymbolicTensor('float32', [1, 3, 'H', 'W'])];
+    const detOutSpec =
+      ocrOpts.detectorKind === 'dbnet'
+        ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
+        : [
+            SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
+            SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
+          ];
+    const detMeta = validateModelSchema(
+      model,
+      `detect_${detBuckets[detBuckets.length - 1]}`,
+      detInSpec,
+      detOutSpec
+    );
+    for (let i = 0; i < detBuckets.length - 1; i++) {
+      validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
+    }
+
+    const built = buildRecognizerSets(model, recBuckets);
+    recSets = built.sets;
+    recC = built.recC;
+    recH = built.recH;
+    recSetByWidth = new Map(recSets.map((s) => [s.width, s]));
+
+    if (recC !== 3) {
+      throw new Error(`OCR: recognizer must take RGB (3 channels), but the model expects ${recC}.`);
+    }
+    charset = buildCharset(ocrOpts.charset);
+    if (charset.length !== built.vocabSize) {
+      throw new Error(
+        `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${built.vocabSize}).`
+      );
+    }
+    // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
+    detExtraChannels = detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
+  } catch (e) {
+    recSets.forEach((s) => {
+      s.tCanvas.dispose();
+      s.tCF.dispose();
+      s.tNorm.dispose();
+      s.tInput.dispose();
+      s.tLogits.dispose();
+    });
+    model.dispose();
+    throw e;
+  }
+
+  // The extractTextBoxes mode matches detectorKind ('craft'/'dbnet').
+  const extractOpts: DetectorExtractConfig =
+    ocrOpts.detectorKind === 'dbnet'
+      ? {
+          mode: 'dbnet',
+          binThreshold: DBNET_BIN_THRESHOLD,
+          boxThreshold: DBNET_BOX_THRESHOLD,
+          unclipRatio: DBNET_UNCLIP_RATIO,
+          minBoxSide: DBNET_MIN_BOX_SIDE,
+          maxCandidates: DBNET_MAX_CANDIDATES,
+          applySigmoid: APPLY_SIGMOID,
+        }
+      : {
+          mode: 'craft',
+          textThreshold: CRAFT_TEXT_THRESHOLD,
+          linkThreshold: CRAFT_LINK_THRESHOLD,
+          lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
+        };
+
+  const dispose = () => {
+    recSets.forEach((s) => {
+      s.tCanvas.dispose();
+      s.tCF.dispose();
+      s.tNorm.dispose();
+      s.tInput.dispose();
+      s.tLogits.dispose();
+    });
+    model.dispose();
+  };
+
+  const runOCRWorklet = (input: ImageBuffer, options?: RunOCROptions): OCRResult => {
+    'worklet';
+    const vertical = options?.vertical ?? false;
+    const tallCropRatio = options?.tallCropRatio ?? TALL_CROP_RATIO;
+    const maxRedetections = options?.maxRedetections ?? MAX_VERTICAL_REDETECTIONS;
+    const { data, width, height, format } = input;
+    const numChannels = FORMAT_CHANNELS[format];
+    // Both detector and recognizer read RGB, so one conversion code serves both.
+    const rgbCode = toRgbCode(format);
+
+    // Detector state, reused for the page pass and the per-box character pass.
+    const detCtx: DetectContext = {
+      model,
+      detectorKind: ocrOpts.detectorKind,
+      detBuckets,
+      format,
+      numChannels,
+      detCode: rgbCode,
+      extractOpts,
+      detExtraChannels,
+    };
+
+    const tInputRaw = tensor('uint8', [height, width, numChannels]);
+    const tRecImage = tensor('uint8', [height, width, recC]);
+    try {
+      tInputRaw.setData(data);
+
+      // ---- detector pass: letterbox -> detect_<S> -> text-box quads (image space) ----
+      const quads = detectQuads(detCtx, tInputRaw, width, height);
+      if (quads.length === 0) {
+        return { detections: [] };
+      }
+
+      // ---- recognizer source: full-res image in RGB ----
+      const recSrc = rgbCode !== null ? cvtColor(tInputRaw, tRecImage, rgbCode) : tInputRaw;
+      const recCtx: RecContext = {
+        model,
+        recSetByWidth,
+        recBuckets,
+        recH,
+        charset,
+      };
+      // The vertical path crops each box from the raw page and re-detects its
+      // characters; `recCode`/`recC` convert a box crop to RGB.
+      const vctx: VerticalContext = {
+        detCtx,
+        rawPage: tInputRaw,
+        recCode: rgbCode,
+        recC,
+        tallCropRatio,
+        redetectBudget: { remaining: maxRedetections },
+      };
+
+      const detections: OCRDetection[] = [];
+      const pushAt =
+        (threshold: number) => (text: string, conf: number, quad: readonly Point[], ms: number) => {
+          if (text.length > 0 && conf >= threshold) {
+            detections.push({ text, confidence: conf, quad, recognizeMs: ms });
+          }
+        };
+      const push = pushAt(dropScore); // flat lines: normal drop-score
+      const pushVertical = pushAt(VERTICAL_DROP_SCORE); // stacked/column: lenient
+
+      // Valid (non-tiny) boxes, ordered TL,TR,BR,BL.
+      const ordered: Point[][] = [];
+      for (const quad of quads) {
+        const o = orderQuad(quad);
+        const s = quadSize(o);
+        if (s.width >= 3 && s.height >= 3) {
+          ordered.push(o);
+        }
+      }
+
+      if (!vertical) {
+        for (const o of ordered) {
+          const boxStart = nowMs();
+          const { text, conf } = recognizeQuad(recCtx, recSrc, o);
+          push(text, conf, o, nowMs() - boxStart);
+        }
+        return { detections };
+      }
+
+      // Additive vertical pass: read x-aligned stacked glyph boxes as one joined
+      // column word; everything else (lines, isolated boxes) reads normally.
+      const { columns, singles } = groupVerticalColumns(ordered);
+      for (const col of columns) {
+        const boxStart = nowMs();
+        const r = recognizeGlyphStrip(recCtx, recSrc, recC, col); // col is top -> bottom
+        if (r) {
+          pushVertical(r.text, r.conf, boundingQuadOf(col), nowMs() - boxStart);
+        }
+      }
+      for (const o of singles) {
+        const boxStart = nowMs();
+        const { text, conf, stacked } = readBoxVertical(recCtx, vctx, recSrc, o, quadSize(o));
+        (stacked ? pushVertical : push)(text, conf, o, nowMs() - boxStart);
+      }
+      return { detections };
+    } finally {
+      tInputRaw.dispose();
+      tRecImage.dispose();
+    }
+  };
+
+  const runOCR = wrapAsync(runOCRWorklet, runtime);
+
+  return { runOCR, runOCRWorklet, dispose };
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
new file mode 100644
index 0000000000..d8884b00c0
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
@@ -0,0 +1,263 @@
+import type { Quad } from '../ops/image';
+import { scalePoint, clamp, type Point } from '../ops/points';
+
+/**
+ * Worklet-safe millisecond clock for in-pipeline profiling. Prefers
+ * `performance.now()` (sub-ms) and falls back to `Date.now()`.
+ * @category Typescript API
+ * @returns The current time in milliseconds.
+ */
+export function nowMs(): number {
+  'worklet';
+  const perf = (globalThis as { performance?: { now?: () => number } }).performance;
+  return perf?.now ? perf.now() : Date.now();
+}
+
+/**
+ * The static input-size buckets a bucketed OCR model exposes. Each PTE ships
+ * per-size methods `detect_<S>` (square `S×S` detector input) and
+ * `recognize_<W>` (fixed recognizer height, width `W`). The client snaps each
+ * input to the closest bucket and calls the matching method. Mirrors each
+ * model's `config.json` `buckets` (today hardcoded in `models.ts`). Both lists
+ * must be ascending.
+ * @category Types
+ */
+export type Buckets = {
+  /** Square detector input sides (e.g. `[640, 960, 1280]`), ascending. */
+  readonly detect: readonly number[];
+  /** Recognizer input widths (e.g. `[160, 320, 480, 640]`), ascending. */
+  readonly recognize: readonly number[];
+};
+
+/**
+ * Snaps an image's longest side *up* to the smallest detector bucket that doesn't
+ * downscale it (clamped to the largest bucket), and returns that square side.
+ * Rounding up, not nearest, is deliberate: downscaling below the detector's
+ * resolution drops text detail and garbles dense lines.
+ * @category Typescript API
+ * @param imgW Original image width.
+ * @param imgH Original image height.
+ * @param buckets The ascending detector side buckets.
+ * @returns The selected square side `S` (call `detect_${S}`).
+ */
+export function snapDetectBucket(imgW: number, imgH: number, buckets: readonly number[]): number {
+  'worklet';
+  const longest = Math.max(imgW, imgH);
+  for (const s of buckets) {
+    if (s >= longest) {
+      return s;
+    }
+  }
+  return buckets[buckets.length - 1]!;
+}
+
+/**
+ * Snaps a desired recognizer content width up to the smallest bucket that fits
+ * it (avoiding horizontal glyph compression), clamped to the largest bucket.
+ * @category Typescript API
+ * @param desiredW The crop/feature content width at the recognizer height.
+ * @param buckets The ascending recognizer width buckets.
+ * @returns The selected width `W` (call `recognize_${W}`).
+ */
+export function snapRecognizeBucket(desiredW: number, buckets: readonly number[]): number {
+  'worklet';
+  for (const w of buckets) {
+    if (w >= desiredW) {
+      return w;
+    }
+  }
+  return buckets[buckets.length - 1]!;
+}
+
+/**
+ * Maps a detector-space quad back into original image pixel coordinates
+ * (reversing the aspect-preserving letterbox), clamping to the image bounds.
+ * @category Typescript API
+ * @param quad The detector-space quad.
+ * @param detW The detector input width the quad is expressed in.
+ * @param detH The detector input height the quad is expressed in.
+ * @param origW Original image width.
+ * @param origH Original image height.
+ * @returns The four quad corners in original image pixels.
+ */
+export function mapQuadToImage(
+  quad: Quad,
+  detW: number,
+  detH: number,
+  origW: number,
+  origH: number
+): Point[] {
+  'worklet';
+  return quad.points.map((p) => {
+    const m = scalePoint(p, {
+      from: { width: detW, height: detH },
+      to: { width: origW, height: origH },
+      resizeMode: 'letterbox',
+    });
+    return { x: clamp(m.x, 0, origW), y: clamp(m.y, 0, origH) };
+  });
+}
+
+/**
+ * Orders four points as TL, TR, BR, BL using x±y extremes (pyimagesearch
+ * convention).
+ * @category Typescript API
+ * @param pts The four unordered quad corners.
+ * @returns The corners ordered TL, TR, BR, BL.
+ */
+export function orderQuad(pts: readonly Point[]): Point[] {
+  'worklet';
+  let tl = 0;
+  let tr = 0;
+  let br = 0;
+  let bl = 0;
+  let minSum = pts[0]!.x + pts[0]!.y;
+  let maxSum = minSum;
+  let minDiff = pts[0]!.y - pts[0]!.x;
+  let maxDiff = minDiff;
+  for (let i = 1; i < 4; i++) {
+    const s = pts[i]!.x + pts[i]!.y;
+    const d = pts[i]!.y - pts[i]!.x;
+    if (s < minSum) {
+      minSum = s;
+      tl = i;
+    }
+    if (s > maxSum) {
+      maxSum = s;
+      br = i;
+    }
+    if (d < minDiff) {
+      minDiff = d;
+      tr = i;
+    }
+    if (d > maxDiff) {
+      maxDiff = d;
+      bl = i;
+    }
+  }
+  return [pts[tl]!, pts[tr]!, pts[br]!, pts[bl]!];
+}
+
+const distance = (a: Point, b: Point): number => {
+  'worklet';
+  return Math.hypot(b.x - a.x, b.y - a.y);
+};
+
+/**
+ * The natural width/height (in pixels) of an ordered TL,TR,BR,BL quad.
+ * @category Typescript API
+ * @param ordered The quad corners ordered TL, TR, BR, BL.
+ * @returns The quad's width and height in pixels.
+ */
+export function quadSize(ordered: readonly Point[]): { width: number; height: number } {
+  'worklet';
+  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
+  const width = Math.max(distance(tl, tr), distance(bl, br));
+  const height = Math.max(distance(tl, bl), distance(tr, br));
+  return { width, height };
+}
+
+/**
+ * Resolves the content width (px) of a recognizer crop: the quad resized to the
+ * recognizer height keeping aspect, clamped to the model's bucket width.
+ * @category Typescript API
+ * @param quadW The quad's natural width.
+ * @param quadH The quad's natural height.
+ * @param recHeight The recognizer input height.
+ * @param bucketWidth The recognizer input (canvas) width.
+ * @returns The clamped content width in pixels.
+ */
+export function contentWidthFor(
+  quadW: number,
+  quadH: number,
+  recHeight: number,
+  bucketWidth: number
+): number {
+  'worklet';
+  const w = Math.round((recHeight * quadW) / Math.max(1, quadH));
+  return clamp(w, 1, bucketWidth);
+}
+
+/**
+ * Flattens an ordered TL,TR,BR,BL quad into the 8-number `[x0,y0,..,x3,y3]`
+ * array expected by `warpQuad`.
+ * @category Typescript API
+ * @param corners The four quad corners (TL, TR, BR, BL).
+ * @returns The eight coordinates `[x0,y0,x1,y1,x2,y2,x3,y3]`.
+ */
+export function flattenQuad(corners: readonly Point[]): number[] {
+  'worklet';
+  // prettier-ignore
+  return [
+    corners[0]!.x, corners[0]!.y, corners[1]!.x, corners[1]!.y,
+    corners[2]!.x, corners[2]!.y, corners[3]!.x, corners[3]!.y,
+  ];
+}
+
+/**
+ * Builds a CTC charset array from a charset string. The first `numSpecials`
+ * indices are reserved special tokens (the CTC blank for CRNN/SVTR); subsequent
+ * entries are the codepoints of `charset` (UTF-8 aware via `Array.from`), so
+ * `charset[id]` decodes logit index `id`.
+ * @category Typescript API
+ * @param charset The ordered character set string.
+ * @param numSpecials Number of reserved low indices (default 1 = CTC blank).
+ * @returns The charset array with `numSpecials` placeholders at the front.
+ */
+export function buildCharset(charset: string | readonly string[], numSpecials = 1): string[] {
+  'worklet';
+  const specials: string[] = [];
+  for (let i = 0; i < numSpecials; i++) {
+    specials.push(`[special${i}]`);
+  }
+  // A string charset is one codepoint per index; an array charset is taken
+  // verbatim (its entries may be multi-codepoint, e.g. ligatures or combining
+  // marks, which `Array.from` would otherwise split).
+  const chars = typeof charset === 'string' ? Array.from(charset) : charset;
+  return [...specials, ...chars];
+}
+
+/**
+ * Greedy CTC decode: skip the reserved special/blank indices (`< numSpecials`)
+ * and consecutive repeats, mapping indices to charset entries.
+ * @category Typescript API
+ * @param indices The per-timestep argmax indices.
+ * @param charset The charset array (specials/blank at the front).
+ * @param numSpecials Number of reserved low indices to drop (default 1).
+ * @returns The decoded text.
+ */
+export function decodeGreedy(indices: number[], charset: string[], numSpecials = 1): string {
+  'worklet';
+  let text = '';
+  let last = -1;
+  for (const idx of indices) {
+    if (idx >= numSpecials && idx !== last && idx < charset.length) {
+      text += charset[idx]!;
+    }
+    last = idx;
+  }
+  return text;
+}
+
+/**
+ * CTC confidence: the mean of the non-special per-timestep max probabilities
+ * (the unified contract softmaxes both heads, so this is comparable across
+ * models). Skips the reserved indices (`< numSpecials`).
+ * @category Typescript API
+ * @param values The per-timestep max probabilities.
+ * @param indices The per-timestep argmax indices.
+ * @param numSpecials Number of reserved low indices to skip (default 1).
+ * @returns The aggregate confidence in `[0, 1]`.
+ */
+export function ctcConfidence(values: number[], indices: number[], numSpecials = 1): number {
+  'worklet';
+  let sum = 0;
+  let count = 0;
+  for (let i = 0; i < indices.length; i++) {
+    if (indices[i]! >= numSpecials) {
+      sum += values[i]!;
+      count++;
+    }
+  }
+  return count === 0 ? 0 : sum / count;
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
new file mode 100644
index 0000000000..5a54d01d71
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
@@ -0,0 +1,261 @@
+import type { WorkletRuntime } from 'react-native-worklets';
+
+import { tensor } from '../../../core/tensor';
+import { loadModel } from '../../../core/model';
+import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
+import { wrapAsync } from '../../../core/runtime';
+
+import type { ImageBuffer } from '../image';
+import { IMAGENET_NORM } from '../../../constants';
+import { gridSample, FORMAT_CHANNELS } from '../ops/image';
+import { createImagePreprocessor } from './preprocessing';
+import { argmaxRange } from './documentHelpers';
+
+// SLANet defaults; overridable per model via SupportingModel (the vocab's eos
+// token id and the structure decoder's max step count).
+const DEFAULT_EOS_TOKEN_ID = 49;
+const DEFAULT_MAX_STEPS = 501; // SLANet max_text_length (500) + 1
+
+/**
+ * Detected page orientation.
+ * @category Types
+ */
+export type Orientation = {
+  /** Degrees the page is rotated clockwise (0/90/180/270). Rotate by `-this` to correct. */
+  readonly rotationCW: 0 | 90 | 180 | 270;
+  readonly confidence: number;
+};
+
+/**
+ * Recognized table structure.
+ * @category Types
+ */
+export type TableStructure = {
+  /** HTML `<tr>/<td>` skeleton (empty cells; fill by aligning OCR text). */
+  readonly html: string;
+  /** Raw SLANet structure token ids (sos/eos stripped). */
+  readonly tokens: number[];
+};
+
+/**
+ * Model configuration for the fused PaddleOCR supporting models (orientation +
+ * dewarp + table structure), all in one PTE.
+ * @category Types
+ */
+export type SupportingModel = {
+  readonly modelPath: string;
+  /** SLANet structure-token vocab (SLANET_STRUCTURE_VOCAB), index = token id. */
+  readonly vocab: readonly string[];
+  /** Token id that terminates AR decoding. Defaults to 49 (SLANet). */
+  readonly eosTokenId?: number;
+  /** Hard cap on AR decode steps. Defaults to 501 (SLANet max length + 1). */
+  readonly maxSteps?: number;
+};
+
+function tokensToHtml(tokens: number[], vocab: readonly string[], eosTokenId: number): string {
+  'worklet';
+  let html = '';
+  for (const t of tokens) {
+    if (t > 0 && t < eosTokenId && t < vocab.length) {
+      html += vocab[t]!;
+    }
+  }
+  return html;
+}
+
+/**
+ * Creates the supporting-models runner (PaddleOCR fused helpers): page
+ * orientation (PP-LCNet), geometric dewarp (UVDoc, applied in TS via the grid),
+ * and table-structure recognition (SLANet_plus AR decode). One PTE, loaded once.
+ * @category Typescript API
+ * @param config Supporting-model path + table structure vocab.
+ * @param runtime Optional worklet runtime thread.
+ * @returns A promise resolving to the three capabilities + disposal controls.
+ */
+export async function createSupporting(
+  config: SupportingModel,
+  runtime?: WorkletRuntime
+): Promise<{
+  dispose: () => void;
+  detectOrientation: (input: ImageBuffer) => Promise<Orientation>;
+  detectOrientationWorklet: (input: ImageBuffer) => Orientation;
+  dewarp: (input: ImageBuffer) => Promise<ImageBuffer>;
+  dewarpWorklet: (input: ImageBuffer) => ImageBuffer;
+  recognizeTable: (input: ImageBuffer) => Promise<TableStructure>;
+  recognizeTableWorklet: (input: ImageBuffer) => TableStructure;
+}> {
+  const { modelPath, vocab } = config;
+  const eosTokenId = config.eosTokenId ?? DEFAULT_EOS_TOKEN_ID;
+  const maxSteps = config.maxSteps ?? DEFAULT_MAX_STEPS;
+  const model = await wrapAsync(loadModel, runtime)(modelPath);
+
+  // --- orientation: image[1,3,224,224] (ImageNet) -> logits[1,4] ---
+  const oriMeta = validateModelSchema(
+    model,
+    'orientation',
+    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+    [SymbolicTensor('float32', [1, 'K'])]
+  );
+  // --- dewarp: image[1,3,712,488] (/255) -> grid[1,2,gH,gW] ---
+  const dewMeta = validateModelSchema(
+    model,
+    'dewarp',
+    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+    [SymbolicTensor('float32', [1, 2, 'gH', 'gW'])]
+  );
+  // --- table backbone + AR decoder ---
+  const encMeta = validateModelSchema(
+    model,
+    'table_encode',
+    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+    [SymbolicTensor('float32', [1, 'C', 'F'])]
+  );
+  const decMeta = validateModelSchema(
+    model,
+    'table_decode_step',
+    [
+      SymbolicTensor('float32', [1, 'C', 'F']),
+      SymbolicTensor('float32', [1, 'H']),
+      SymbolicTensor('float32', [1, 'V']),
+    ],
+    [SymbolicTensor('float32', [1, 'V']), SymbolicTensor('float32', [1, 'H'])]
+  );
+
+  const oriShape = oriMeta.inputTensorMeta[0]!.shape;
+  const oriOutLen = oriMeta.outputTensorMeta[0]!.shape[1]!;
+  const dewShape = dewMeta.inputTensorMeta[0]!.shape;
+  const gridShape = dewMeta.outputTensorMeta[0]!.shape; // [1,2,gH,gW]
+  const tabShape = encMeta.inputTensorMeta[0]!.shape;
+  const featShape = encMeta.outputTensorMeta[0]!.shape;
+  const hidShape = decMeta.outputTensorMeta[1]!.shape;
+  const probShape = decMeta.outputTensorMeta[0]!.shape;
+  const hidLen = hidShape[1]!;
+  const vocabLen = probShape[1]!;
+
+  if (vocabLen !== vocab.length) {
+    throw new Error(
+      `Supporting: vocab length (${vocab.length}) must match the model's token dim (${vocabLen}).`
+    );
+  }
+
+  const oriPre = createImagePreprocessor(
+    {
+      resizeMode: 'stretch',
+      interpolation: 'linear',
+      alpha: IMAGENET_NORM.alpha,
+      beta: IMAGENET_NORM.beta,
+    },
+    oriShape
+  );
+  const dewPre = createImagePreprocessor(
+    { resizeMode: 'stretch', interpolation: 'linear', alpha: 1 / 255, beta: 0 },
+    dewShape
+  );
+  const tabPre = createImagePreprocessor(
+    {
+      resizeMode: 'stretch',
+      interpolation: 'linear',
+      alpha: IMAGENET_NORM.alpha,
+      beta: IMAGENET_NORM.beta,
+    },
+    tabShape
+  );
+
+  const tensors = [
+    tensor('float32', oriMeta.outputTensorMeta[0]!.shape), // tOri
+    tensor('float32', gridShape), // tGrid
+    tensor('float32', featShape), // tFeatures
+    tensor('float32', hidShape), // tHidden
+    tensor('float32', probShape), // tOnehot
+    tensor('float32', probShape), // tProbs
+    tensor('float32', hidShape), // tNewHidden
+  ] as const;
+  const [tOri, tGrid, tFeatures, tHidden, tOnehot, tProbs, tNewHidden] = tensors;
+  const oriBuf = new Float32Array(oriOutLen);
+  const zeroHidden = new Float32Array(hidLen);
+  const zeroVocab = new Float32Array(vocabLen);
+  const onehotBuf = new Float32Array(vocabLen);
+  const probsBuf = new Float32Array(vocabLen);
+
+  const dispose = () => {
+    oriPre.dispose();
+    dewPre.dispose();
+    tabPre.dispose();
+    tensors.forEach((t) => t.dispose());
+    model.dispose();
+  };
+
+  const detectOrientationWorklet = (input: ImageBuffer): Orientation => {
+    'worklet';
+    const tInput = oriPre.process(input);
+    model.execute('orientation', [tInput], [tOri]);
+    tOri.getData(oriBuf);
+    const cls = argmaxRange(oriBuf, 0, oriOutLen);
+    const best = oriBuf[cls]!;
+    let sumExp = 0;
+    for (let i = 0; i < oriOutLen; i++) {
+      sumExp += Math.exp(oriBuf[i]! - best);
+    }
+    return { rotationCW: ((cls % 4) * 90) as 0 | 90 | 180 | 270, confidence: 1 / sumExp };
+  };
+
+  const dewarpWorklet = (input: ImageBuffer): ImageBuffer => {
+    'worklet';
+    const tInput = dewPre.process(input);
+    model.execute('dewarp', [tInput], [tGrid]);
+    // Apply the sampling field to the full-res page natively (cv::remap). The
+    // page-sized src/dst tensors depend on the input size, so allocate per call.
+    const ch = FORMAT_CHANNELS[input.format];
+    const tSrc = tensor('uint8', [input.height, input.width, ch]);
+    const tDst = tensor('uint8', [input.height, input.width, ch]);
+    try {
+      tSrc.setData(input.data);
+      gridSample(tSrc, tGrid, tDst);
+      const out = new Uint8Array(input.width * input.height * ch);
+      tDst.getData(out);
+      return {
+        data: out,
+        width: input.width,
+        height: input.height,
+        format: input.format,
+        layout: input.layout,
+      };
+    } finally {
+      tSrc.dispose();
+      tDst.dispose();
+    }
+  };
+
+  const recognizeTableWorklet = (input: ImageBuffer): TableStructure => {
+    'worklet';
+    const tInput = tabPre.process(input);
+    model.execute('table_encode', [tInput], [tFeatures]);
+    tHidden.setData(zeroHidden);
+    tOnehot.setData(zeroVocab);
+    const tokens: number[] = [];
+    for (let step = 0; step < maxSteps; step++) {
+      model.execute('table_decode_step', [tFeatures, tHidden, tOnehot], [tProbs, tNewHidden]);
+      tProbs.getData(probsBuf);
+      const tok = argmaxRange(probsBuf, 0, vocabLen);
+      tokens.push(tok);
+      if (tok === eosTokenId) {
+        break;
+      }
+      tNewHidden.copyTo(tHidden);
+      onehotBuf.fill(0);
+      onehotBuf[tok] = 1;
+      tOnehot.setData(onehotBuf);
+    }
+    return { html: tokensToHtml(tokens, vocab, eosTokenId), tokens };
+  };
+
+  return {
+    dispose,
+    detectOrientation: wrapAsync(detectOrientationWorklet, runtime),
+    detectOrientationWorklet,
+    dewarp: wrapAsync(dewarpWorklet, runtime),
+    dewarpWorklet,
+    recognizeTable: wrapAsync(recognizeTableWorklet, runtime),
+    recognizeTableWorklet,
+  };
+}
diff --git a/packages/react-native-executorch/src/hooks/useDocumentOCR.ts b/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
new file mode 100644
index 0000000000..a48c11d6a4
--- /dev/null
+++ b/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
@@ -0,0 +1,59 @@
+import { useModel } from './useModel';
+import { useResourceDownload } from './useResourceDownload';
+import { createDocumentOCR, type DocumentOCRModel } from '../extensions/cv/tasks/documentOCR';
+
+/**
+ * React hook for the document OCR pipeline: OCR + optional layout detection +
+ * optional supporting (orientation/dewarp/table), assembled into reading-ordered
+ * blocks. Downloads/compiles all enabled models, tracks progress and errors, and
+ * cleans up native memory on unmount or config change.
+ * @category Hooks
+ * @typeParam L The type representing the layout region class labels.
+ * @param config OCR model + optional layout/supporting models + flags.
+ * @param options Hook options.
+ * @param options.preventLoad If true, prevents downloading and compiling the models.
+ * @returns Loading state, error, download progress, and the document run functions.
+ */
+export function useDocumentOCR<L>(
+  config: DocumentOCRModel<L>,
+  options?: { preventLoad?: boolean }
+) {
+  const ocrDl = useResourceDownload(config.ocr.modelPath, options?.preventLoad);
+  const layoutDl = useResourceDownload(config.layout?.modelPath, options?.preventLoad);
+  const supDl = useResourceDownload(config.supporting?.modelPath, options?.preventLoad);
+
+  const ready =
+    !!ocrDl.localPath &&
+    (!config.layout || !!layoutDl.localPath) &&
+    (!config.supporting || !!supDl.localPath);
+  const localConfig: DocumentOCRModel<L> | null = ready
+    ? {
+        ...config,
+        ocr: { ...config.ocr, modelPath: ocrDl.localPath! },
+        ...(config.layout ? { layout: { ...config.layout, modelPath: layoutDl.localPath! } } : {}),
+        ...(config.supporting
+          ? { supporting: { ...config.supporting, modelPath: supDl.localPath! } }
+          : {}),
+      }
+    : null;
+
+  const { model, error } = useModel(createDocumentOCR<L>, localConfig, [
+    ocrDl.localPath,
+    layoutDl.localPath,
+    supDl.localPath,
+  ]);
+
+  // Overall progress is the slowest of the enabled downloads, so it can't read
+  // 100% while a second/third model is still fetching.
+  const progresses = [ocrDl.downloadProgress];
+  if (config.layout) progresses.push(layoutDl.downloadProgress);
+  if (config.supporting) progresses.push(supDl.downloadProgress);
+
+  return {
+    isReady: !!model,
+    error: ocrDl.downloadError || layoutDl.downloadError || supDl.downloadError || error,
+    downloadProgress: Math.min(...progresses),
+    runDocumentOCR: model?.runDocumentOCR,
+    runDocumentOCRWorklet: model?.runDocumentOCRWorklet,
+  };
+}
diff --git a/packages/react-native-executorch/src/hooks/useOCR.ts b/packages/react-native-executorch/src/hooks/useOCR.ts
new file mode 100644
index 0000000000..0699f4c1f1
--- /dev/null
+++ b/packages/react-native-executorch/src/hooks/useOCR.ts
@@ -0,0 +1,39 @@
+import { createOCR, type OCRModel } from '../extensions/cv/tasks/ocr';
+import { useResourceDownload } from './useResourceDownload';
+import { useModel } from './useModel';
+
+/**
+ * React hook for running the unified OCR pipeline (EasyOCR / PaddleOCR).
+ *
+ * Downloads the fused PTE, instantiates the OCR task runner, and manages its
+ * lifetime. Heavy work runs on a worklet thread; the returned `runOCR` resolves
+ * with the recognized text regions.
+ * @category Hooks
+ * @param config OCR model configuration (one fused PTE + flat options). Use a
+ * preset from `models.ocr.*`.
+ * @param options Optional flags. `preventLoad` defers downloading/compiling the
+ * model until set to `false`.
+ * @returns Readiness flags, download progress, and the `runOCR` /
+ * `runOCRWorklet` runners.
+ */
+export function useOCR(config: OCRModel, options?: { preventLoad?: boolean }) {
+  const { localPath, downloadProgress, downloadError } = useResourceDownload(
+    config.modelPath,
+    options?.preventLoad
+  );
+
+  const { model, error } = useModel(
+    createOCR,
+    localPath ? { ...config, modelPath: localPath } : null,
+    [localPath]
+  );
+
+  return {
+    isReady: !!model,
+    error: downloadError || error,
+    downloadProgress,
+    localPath,
+    runOCR: model?.runOCR,
+    runOCRWorklet: model?.runOCRWorklet,
+  };
+}
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 00557be78b..5fdfd6b7da 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -6,6 +6,10 @@ export * from './hooks/useInstanceSegmenter';
 export * from './hooks/useKeypointDetector';
 export * from './hooks/useObjectDetector';
 export * from './hooks/useTokenizer';
+export * from './hooks/useOCR';
+// Layout + supporting (orientation/dewarp/table) are internal to the document
+// pipeline — consumed by useDocumentOCR, not exposed as standalone hooks.
+export * from './hooks/useDocumentOCR';
 export * from './hooks/useResourceDownload';
 export * from './hooks/useModel';
 
@@ -21,6 +25,9 @@ export * from './extensions/cv/tasks/instanceSegmentation';
 export * from './extensions/cv/tasks/keypointDetection';
 export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/nlp/tasks/tokenization';
+export * from './extensions/cv/tasks/ocr';
+export * from './extensions/cv/tasks/supporting';
+export * from './extensions/cv/tasks/documentOCR';
 
 // Core primitives — for library builders and power users
 export { tensor } from './core/tensor';
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 3397288e23..1c3e5c1ce2 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -4,6 +4,8 @@ import type { StyleTransferModel } from './extensions/cv/tasks/styleTransfer';
 import type { SemanticSegmentationModel } from './extensions/cv/tasks/semanticSegmentation';
 import type { KeypointDetectorModel } from './extensions/cv/tasks/keypointDetection';
 import type { InstanceSegmenterModel } from './extensions/cv/tasks/instanceSegmentation';
+import type { OCRModel, OCROptions } from './extensions/cv/tasks/ocr';
+import type { SupportingModel } from './extensions/cv/tasks/supporting';
 import {
   IMAGENET_NORM,
   IMAGENET1K_LABELS,
@@ -12,12 +14,17 @@ import {
   COCO_CLASSES_YOLO,
   BLAZEFACE_LANDMARKS,
   COCO_LANDMARKS,
+  DOC_LAYOUT_LABELS,
+  SLANET_STRUCTURE_VOCAB,
+  alphabets,
+  PPOCR_SYMBOLS,
   type ImageNet1KLabel,
   type PascalVocLabel,
   type CocoClass,
   type CocoClassYolo,
   type BlazeFaceLandmark,
   type CocoLandmark,
+  type DocLayoutLabel,
 } from './constants';
 
 const BASE_URL = 'https://huggingface.co/software-mansion/react-native-executorch';
@@ -532,6 +539,100 @@ const YOLO26_XLARGE_SEG_640_XNNPACK_FP32: InstanceSegmenterModel<'xyxy', CocoCla
 // Tokenizers
 // =============================================================================
 const ALL_MINILM_L6_V2_TOKENIZER = `${BASE_URL}-all-MiniLM-L6-v2/${VERSION_TAG}/tokenizer.json`;
+// OCR
+// =============================================================================
+// EasyOCR (CRAFT + CRNN) and PaddleOCR (DBNet + SVTR). The detector/recognizer
+// profile (normalization, color, padding, CTC head, confidence) is derived from
+// `detectorKind` inside the task, so each model only declares its architecture,
+// input-size `buckets`, and `charset`. `charset` for EasyOCR is set per language.
+const EASYOCR_OPTS: OCROptions = {
+  detectorKind: 'craft', // CRAFT: text + affinity heatmaps grouped into lines
+  charset: alphabets.english, // overridden per language
+  buckets: { detect: [800, 1280], recognize: [64, 128, 256, 512] },
+};
+
+const PADDLE_PPOCRV6_OPTS: OCROptions = {
+  detectorKind: 'dbnet', // DBNet: binary probability map contoured into boxes
+  charset: PPOCR_SYMBOLS,
+  buckets: { detect: [640, 960, 1280], recognize: [160, 320, 480, 640, 1280] },
+};
+
+// OCR models live in dedicated per-architecture HF repos, pinned to `main`. Each
+// PTE is bucketed: it ships per-size `detect_<S>` / `recognize_<W>` methods and
+// the pipeline snaps each image to the closest bucket. Each repo also hosts a
+// `config.json` (detectorKind, buckets, dropScore, charsetUrl) and `charset.txt`;
+// for now those stay bundled in the family OCROptions above / constants rather
+// than being fetched.
+const OCR_REVISION = 'resolve/main';
+const EASYOCR_REPO = `${BASE_URL}-EasyOCR/${OCR_REVISION}`;
+const PPOCRV6_REPO = `${BASE_URL}-PP-OCRv6/${OCR_REVISION}`;
+const PP_DOCLAYOUT_REPO = `${BASE_URL}-PP-DocLayoutV3/${OCR_REVISION}`;
+const PADDLE_HELPERS_REPO = `${BASE_URL}-PaddleHelpers/${OCR_REVISION}`;
+
+type OCRBackend = 'xnnpack' | 'coreml' | 'vulkan';
+
+// EasyOCR hosts one subdir per language (`<lang>/EasyOCR_<lang>_<backend>.pte`);
+// the detector + pipeline are shared, only the recognizer head + charset change.
+const easyocr = (lang: string, backend: OCRBackend, charset: string): OCRModel => ({
+  modelPath: `${EASYOCR_REPO}/${lang}/EasyOCR_${lang}_${backend}.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset },
+});
+const paddle = (backend: OCRBackend): OCRModel => ({
+  modelPath: `${PPOCRV6_REPO}/PP-OCRv6_${backend}.pte`,
+  ocrOpts: PADDLE_PPOCRV6_OPTS,
+});
+
+// EasyOCR ships 8 language recognizers (XNNPACK + CoreML + Vulkan each).
+// NOTE: per the export notes, zh_sim recognize is int8-fragile on XNNPACK/Vulkan
+// (large vocab) — prefer CoreML there.
+const easyLang = (lang: string, charset: string) => ({
+  XNNPACK: easyocr(lang, 'xnnpack', charset),
+  COREML: easyocr(lang, 'coreml', charset),
+  VULKAN: easyocr(lang, 'vulkan', charset),
+});
+
+const PADDLE_PPOCRV6_XNNPACK = paddle('xnnpack');
+const PADDLE_PPOCRV6_VULKAN = paddle('vulkan');
+const PADDLE_PPOCRV6_COREML = paddle('coreml');
+
+// =============================================================================
+// Document layout — PP-DocLayoutV3 (RT-DETR, 25 region classes, fixed 800x800).
+// =============================================================================
+// PP-DocLayoutV3 speaks the objectDetection contract (the RT-DETR decode is baked
+// in-graph): RGB 800x800, x/255 with NO mean/std, stretch. NMS is disabled
+// (iouThreshold 1) because RT-DETR is set-prediction and standard NMS would merge
+// legitimately nested regions (e.g. a table inside a figure).
+const PP_DOCLAYOUT_OPTS = {
+  labels: DOC_LAYOUT_LABELS,
+  boxFormat: 'xyxy' as const,
+  resizeMode: 'stretch' as const,
+  interpolation: 'linear' as const,
+  alpha: 1 / 255.0,
+  beta: 0.0,
+  defaultConfidenceThreshold: 0.3,
+  defaultIouThreshold: 1.0,
+};
+const layout = (backend: OCRBackend): ObjectDetectorModel<'xyxy', DocLayoutLabel> => ({
+  modelPath: `${PP_DOCLAYOUT_REPO}/PP-DocLayoutV3_${backend}.pte`,
+  opts: PP_DOCLAYOUT_OPTS,
+});
+const PP_DOCLAYOUT_XNNPACK = layout('xnnpack');
+const PP_DOCLAYOUT_VULKAN = layout('vulkan');
+const PP_DOCLAYOUT_COREML = layout('coreml');
+
+// =============================================================================
+// Table structure — SLANet_plus (in the fused supporting PTE: table_encode +
+// table_decode_step). Also carries orientation/dewarp methods (same file).
+// =============================================================================
+const supporting = (backend: OCRBackend): SupportingModel => ({
+  modelPath: `${PADDLE_HELPERS_REPO}/PaddleHelpers_${backend}.pte`,
+  vocab: SLANET_STRUCTURE_VOCAB,
+  eosTokenId: 49, // SLANet '</s>' token id
+  maxSteps: 501, // SLANet max_text_length (500) + 1
+});
+const PP_TABLE_XNNPACK = supporting('xnnpack');
+const PP_TABLE_VULKAN = supporting('vulkan');
+const PP_TABLE_COREML = supporting('coreml');
 
 /**
  * Registry of pre-configured ExecuTorch models.
@@ -737,4 +838,37 @@ export const models = {
   tokenizer: {
     ALL_MINILM_L6_V2: ALL_MINILM_L6_V2_TOKENIZER,
   },
+  ocr: {
+    EASYOCR: {
+      ENGLISH: easyLang('english', alphabets.english),
+      CYRILLIC: easyLang('cyrillic', alphabets.cyrillic),
+      LATIN: easyLang('latin', alphabets.latin),
+      JAPANESE: easyLang('japanese', alphabets.japanese),
+      ZH_SIM: easyLang('zh_sim', alphabets.zh_sim),
+      KOREAN: easyLang('korean', alphabets.korean),
+      TELUGU: easyLang('telugu', alphabets.telugu),
+      KANNADA: easyLang('kannada', alphabets.kannada),
+    },
+    PADDLE: {
+      PPOCRV6_SMALL: {
+        XNNPACK: PADDLE_PPOCRV6_XNNPACK,
+        VULKAN: PADDLE_PPOCRV6_VULKAN,
+        COREML: PADDLE_PPOCRV6_COREML,
+      },
+    },
+  },
+  layoutDetection: {
+    PP_DOCLAYOUT: {
+      XNNPACK: PP_DOCLAYOUT_XNNPACK,
+      VULKAN: PP_DOCLAYOUT_VULKAN,
+      COREML: PP_DOCLAYOUT_COREML,
+    },
+  },
+  supporting: {
+    PP_SUPPORTING: {
+      XNNPACK: PP_TABLE_XNNPACK,
+      VULKAN: PP_TABLE_VULKAN,
+      COREML: PP_TABLE_COREML,
+    },
+  },
 };
diff --git a/packages/react-native-executorch/src/ocrSymbols.ts b/packages/react-native-executorch/src/ocrSymbols.ts
new file mode 100644
index 0000000000..64c7c40c14
--- /dev/null
+++ b/packages/react-native-executorch/src/ocrSymbols.ts
@@ -0,0 +1,161 @@
+/* eslint-disable @cspell/spellchecker */
+/* eslint-disable camelcase */
+export const alphabets = {
+  cyrillic:
+    '0123456789!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~ €₽ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyzАБВГДЕЁЖЗИЙКЛМНОПРСТУФХЦЧШЩЪЫЬЭЮЯабвгдеёжзийклмнопрстуфхцчшщъыьэюяЂђЃѓЄєІіЇїЈјЉљЊњЋћЌќЎўЏџҐґҒғҚқҮүҲҳҶҷӀӏӢӣӨөӮӯ',
+  english:
+    '0123456789!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~ €ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz',
+  latin:
+    ' !"#$%&\'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~ªÀÁÂÃÄÅÆÇÈÉÊËÌÍÎÏÐÑÒÓÔÕÖØÙÚÛÜÝÞßàáâãäåæçèéêëìíîïðñòóôõöøùúûüýþÿĀāĂ㥹ĆćČčĎďĐđĒēĖėĘęĚěĞğĨĩĪīĮįİıĶķĹĺĻļĽľŁłŃńŅņŇňŒœŔŕŘřŚśŞşŠšŤťŨũŪūŮůŲųŸŹźŻżŽžƏƠơƯưȘșȚțə̇ḌḍḶḷṀṁṂṃṄṅṆṇṬṭẠạẢảẤấẦầẨẩẪẫẬậẮắẰằẲẳẴẵẶặẸẹẺẻẼẽẾếỀềỂểỄễỆệỈỉỊịỌọỎỏỐốỒồỔổỖỗỘộỚớỜờỞởỠỡỢợỤụỦủỨứỪừỬửỮữỰựỲỳỴỵỶỷỸỹ€',
+  japanese:
+    ' !"#$%&\'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~‥…※♪、。々〈〉《》「」『』【】〔〕あぃいうぇえおかがきぎくぐけげこごさざしじすずせぜそぞただちぢっつづてでとどなにぬねのはばぱひびぴふぶぷへべぺほぼぽまみむめもゃやゅゆょよらりるれろわをんァアィイゥウェエォオカガキギクグケゲコゴサザシジスズセゼソゾタダチヂッツヅテデトドナニヌネノハバパヒビピフブプヘベペホボポマミムメモャヤュユョヨラリルレロワヱヲンヴヶ・ー一丁七万丈三上下不与丑世丘丙丞両並中串丸丹主乃久之乗乙九也乱乳乾亀了予争事二互五井亘亜亡交亥亨享京亭亮人仁今介仏仕他付仙代令以仮仰仲件任企伊伍伎伏休会伝伯伴伶伸似伽位低住佐佑体何余作佳使例侍侑供依価便係俊俗保信俣修俵倉個倍倒倖候借値倫倭倶偉偏健側偵偽傍傑備債傷働像僧儀優允元兄充先光克免兎児党入全八公六共兵其具典兼内円再写冠冨冬冴冶冷凍凡処凪出刀刃分切刈刑列初判別利制刷券刺刻則削前剛剣剤剥副割創劇力功加助労効勅勇勉動勘務勝勢勤勧勲包化北匠匡区医十千升午半卒卓協南単博占卯印危即卵卸厚原厨厳去参又叉及友双反収取受叡口古句只叫召可台史右叶号司各合吉同名后吏吐向君吟否吸吹吾呂呉告呑周味呼命和咲哀品哉員哲唄唐唯唱商問啓善喜喬喰営嗣嘉噌器四回因団囲図固国國園土圧在圭地坂均坊坐坪垂型垢垣埋城埜域執基埼堀堂堅堤堰報場堺塔塗塚塩塵境墓増墨墳壁壇壊士壬壮声壱売壷変夏夕外多夜夢大天太夫央失夷奇奈奉奏契奥奨女奴好如妃妙妹妻姉始姓委姥姫姿威娘婆婚婦嫌嬉子孔字存孝孟季孤学孫宅宇守安完宏宗官宙定宜宝実客宣室宥宮宰害家容宿寂寄寅密富寒寛寝察寧審寸寺対寿封専射将尊尋導小少尚尭就尺尻尼尽尾尿局居屈屋展属層屯山岐岡岩岬岱岳岸峠峡峨峯峰島峻崇崎崩嵐嵩嵯嶋嶺巌川州巡巣工左巧巨差己巳巴巻巾市布帆希帝師席帯帰帳常帽幅幌幕幡幣干平年幸幹幽幾庁広庄床序底店府度座庫庭庵康庸廃廉廣延建廻弁式弐弓引弘弟弥弦弱張強弾当形彦彩彪彫彬彰影役彼往征径待律後徒従得御復微徳徹心必忌忍志応忠快念怒怜思急性恋恐恒恩恭息恵悌悟悠患悦悪悲情惇惑惟惣想意愚愛感慈態慎慣慧慶憂憲憶懐懸戎成我戒戦戯戸戻房所扇手才打払扶承技投抗折抜抱押担拓拝拡拳拾持指挙振捕捨捷掃排掘掛採探接推掻提揖揚換揮援揺損摂摘摩摺撃撫播撮操擦擬支改攻放政故敏救敗教敢散敦敬数整敵敷文斉斎斐斑斗料斜斤断斯新方於施旅旋族旗日旦旧旨早旬旭旺昂昆昇昌明易星映春昭是昼時晃晋晩普景晴晶智暁暖暗暢暦暮暴曇曙曜曲曳更書曹曽曾替最月有朋服朔朗望朝期木未末本札朱朴杉李杏材村杖杜束条来杭東杵松板析枕林枚果枝枯架柄柊柏柑染柔柚柱柳柴査柿栃栄栖栗校株核根格桂桃案桐桑桜桝桧桶梁梅梓梢梨梯械梶棄棒棚棟森椋植椎検椿楊楓楠楢業楯極楼楽榊榎榛構槌様槙槻樋標模権横樫樹樺樽橋橘機檀櫛欠次欣欧欲欽歌歓止正此武歩歯歳歴死殊残殖段殺殻殿毅母毎毒比毛氏民気水氷永汀汁求汐汗汚江池汰汲決沈沓沖沙沢河油治沼泉泊法波泣泥注泰洋洗洞津洪洲活派流浄浅浜浦浩浪浮浴海消涌涙液涼淀淑淡深淳淵混添清済渉渋渓渕渚減渡渥温測港湊湖湧湯湾湿満源準溜溝溶滅滋滑滝漁漆漏演漢漬潔潜潟潤潮潰澄澤激濃濱瀧瀬灘火灯灰災炉炎炭点為烈烏無焦然焼煙照煮熊熟熱燃燈燕燦燭爆爪父爽爾片版牛牟牧物特犬犯状狂狐狗狩独狭狼猛猪猫献猿獄獅獣玄率玉王玖玲珍珠現球理琉琢琳琴瑚瑛瑞瑠瑳璃環瓜瓦瓶甘甚生産用甫田由甲申男町画界畑畔留畜畝畠略番異畳疾病症痛療発登白百的皆皇皮皿盆益盗盛盟監盤目盲直相省眉看県眞真眠眼着督睦瞬瞳矢知矩短石砂研砲破硫硬碑碧碩確磁磐磨磯礁示礼社祇祈祉祐祖祝神祢祥票祭禁禄禅禎福禰秀私秋科秘秦秩称移稀程税稔稗稚種稲穂積穏穴究空突窓窪立竜章童竪端競竹笑笛笠符第笹筆等筋筑筒答策箇箕算管箱箸節範築篠篤篭簡簾籍米粉粒粕粗粟粥精糖糞糠糸系紀約紅紋納純紗紘紙級素紡索紫細紳紹紺終組経結絡絢給統絵絶絹継続綜維綱網綾綿緋総緑緒線締編緩練縁縄縦縫縮績繁織繰罪置羅羊美群義羽翁習翔翠翼耀老考者耐耕耳耶聖聞聡聴職肇肉肌肝股肥肩育肺背胞胡胤胸能脂脇脈脚脱脳腐腕腫腰腸腹膜膳臣臥臨自臭至致臼興舌舎舘舛舜舞舟航般船艦良色艶芋芙芝芥芦花芳芸芹芽苅苑苔苗若苦苫英茂茄茅茉茜茨茶草荒荘荷荻莉菅菊菌菓菖菜華菱萌萩萱落葉葛葦葵蒔蒲蒸蒼蓋蓑蓬蓮蔦蔭蔵蕗薄薩薫薬薮藁藍藤藻蘇蘭虎虚虫虹虻蚊蛇蛍蛭蜂蜜蝦蝶融螺蟹蟻血衆行術街衛衝衡衣表袋袖被裁裂装裏裕補裟裸製複西要覆覇見規視覚覧親観角解触言計訓託記訪設許訳訴診証評詞詠試詩詰話誉誌認誓誘語誠誤説読課調談請諏論諭諸謙講謝謹識警議譲護讃谷豆豊豚象豪貝貞負財貢貧貨販貫責貯貴買貸費賀賃資賞賢質赤赦走起超越足跡路跳踏身車軌軍軒軟転軸軽載輔輝輪輸辛辞辰農辺辻込迎近返迦迫述迷追退送逃逆透途通速造逢連週進逸遅遊運過道達違遠遣遥適選遺遼避邑那邦邪郁郎郡部郭郵郷都配酒酔酢酸醍醐采釈里重野量金釘釜針釣鈴鉄鉛鉢鉱鉾銀銃銅銘銭鋭鋼錦録鍋鍛鍬鍵鎌鎖鎮鏡鐘鑑長門閉開閑間関閣闇闘阪防阿陀附降限院陣除陰陳陵陶陸険陽隅隆隈隊階随隔際障隠隣隼雀雁雄雅集雑雛離難雨雪雲零雷電震霊霜霞霧露青靖静非面革鞍鞠韓音響頂頃項順須預頓領頭頼題額顔顕願類風飛食飯飲飼飽飾餅養館首香馨馬駄駅駆駐駒駿騎験骨高髪鬼魁魂魅魔魚鮎鮫鮮鯉鯨鳥鳩鳳鳴鴨鴻鵜鶏鶴鷲鷹鷺鹿麓麗麦麻麿黄黒黙鼓鼠鼻齢龍＊',
+  zh_sim:
+    ' !"#$%&\'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~。〈〉《》「」『』一丁七万丈三上下不与丐丑专且丕世丘丙业丛东丝丞丢两严丧个丫中丰串临丸丹为主丽举乃久么义之乌乍乎乏乐乒乓乔乖乘乙乜九乞也习乡书乩买乱乳乾了予争事二亍于亏云互亓五井亘亚些亟亡亢交亥亦产亨亩享京亭亮亲亳亵人亿什仁仂仃仄仅仆仇仉今介仍从仑仓仔仕他仗付仙仞仟仡代令以仨仪仫们仰仲仳仵件价任份仿企伉伊伍伎伏伐休众优伙会伛伞伟传伢伤伦伧伪伫伯估伴伶伸伺似伽佃但位低住佐佑体何佗佘余佚佛作佝佞佟你佣佤佥佩佬佯佰佳佴佶佻佼佾使侃侄侈侉例侍侏侑侔侗供依侠侣侥侦侧侨侩侪侬侮侯侵便促俄俅俊俎俏俐俑俗俘俚俜保俞俟信俣俦俨俩俪俭修俯俱俳俸俺俾倌倍倏倒倔倘候倚倜借倡倥倦倨倩倪倬倭债值倾偃假偈偌偎偏偕做停健偬偶偷偻偾偿傀傅傈傍傣傥傧储傩催傲傻像僖僚僦僧僬僭僮僳僵僻儆儇儋儒儡儿兀允元兄充兆先光克免兑兔兕兖党兜兢入全八公六兮兰共关兴兵其具典兹养兼兽冀冁内冈冉册再冒冕冗写军农冠冢冤冥冬冯冰冱冲决况冶冷冻冼冽净凄准凇凉凋凌减凑凛凝几凡凤凫凭凯凰凳凶凸凹出击凼函凿刀刁刃分切刈刊刍刎刑划刖列刘则刚创初删判刨利别刭刮到刳制刷券刹刺刻刽刿剀剁剂剃削剌前剐剑剔剖剜剞剡剥剧剩剪副割剽剿劁劂劈劐劓力劝办功加务劢劣动助努劫劬劭励劲劳劾势勃勇勉勋勐勒勖勘募勤勰勺勾勿匀包匆匈匍匏匐匕化北匙匝匠匡匣匦匪匮匹区医匾匿十千卅升午卉半华协卑卒卓单卖南博卜卞卟占卡卢卣卤卦卧卫卮卯印危即却卵卷卸卺卿厂厄厅历厉压厌厍厕厘厚厝原厢厣厥厦厨厩厮去县叁参又叉及友双反发叔取受变叙叛叟叠口古句另叨叩只叫召叭叮可台叱史右叵叶号司叹叻叼叽吁吃各吆合吉吊同名后吏吐向吒吓吕吗君吝吞吟吠吡吣否吧吨吩含听吭吮启吱吲吴吵吸吹吻吼吾呀呃呆呈告呋呐呓呔呕呖呗员呙呛呜呢呤呦周呱呲味呵呶呷呸呻呼命咀咂咄咆咋和咎咏咐咒咔咕咖咙咚咛咝咣咤咦咧咨咩咪咫咬咭咯咱咳咴咸咻咽咿哀品哂哄哆哇哈哉哌响哎哏哐哑哓哔哕哗哙哚哝哞哟哥哦哧哨哩哪哭哮哲哳哺哼哽哿唁唆唇唉唏唐唑唔唛唠唢唣唤唧唪唬售唯唱唳唷唼唾唿啁啃啄商啉啊啐啕啖啜啡啤啥啦啧啪啬啭啮啵啶啷啸啻啼啾喀喁喂喃善喇喈喉喊喋喏喑喔喘喙喜喝喟喧喱喳喵喷喹喻喽喾嗄嗅嗉嗌嗍嗑嗒嗓嗔嗖嗜嗝嗟嗡嗣嗤嗥嗦嗨嗪嗫嗬嗯嗲嗳嗵嗷嗽嗾嘀嘁嘈嘉嘌嘎嘏嘘嘛嘞嘟嘣嘤嘧嘬嘭嘱嘲嘴嘶嘹嘻嘿噌噍噎噔噗噙噜噢噤器噩噪噫噬噱噶噻噼嚅嚆嚎嚏嚓嚣嚯嚷嚼囊囔囚四囝回囟因囡团囤囫园困囱围囵囹固国图囿圃圄圆圈圉圊圜土圣在圩圪圬圭圮圯地圳圹场圻圾址坂均坊坌坍坎坏坐坑块坚坛坜坝坞坟坠坡坤坦坨坩坪坫坭坯坳坷坻坼垂垃垄垅垆型垌垒垓垛垠垡垢垣垤垦垧垩垫垭垮垲垸埂埃埋城埏埒埔埕埘埙埚埝域埠埤埭埯埴埸培基埽堂堆堇堋堍堑堕堙堞堠堡堤堪堰堵塄塌塍塑塔塘塞塥填塬塾墀墁境墅墉墒墓墙增墟墨墩墼壁壅壑壕壤士壬壮声壳壶壹处备复夏夔夕外夙多夜够夤夥大天太夫夭央夯失头夷夸夹夺夼奁奂奄奇奈奉奋奎奏契奔奕奖套奘奚奠奢奥女奴奶奸她好妁如妃妄妆妇妈妊妍妒妓妖妗妙妞妣妤妥妨妩妪妫妮妯妲妹妻妾姆姊始姐姑姒姓委姗姘姚姜姝姣姥姨姬姹姻姿威娃娄娅娆娇娈娉娌娑娓娘娜娟娠娣娥娩娱娲娴娶娼婀婆婉婊婕婚婢婧婪婴婵婶婷婺婿媒媚媛媪媲媳媵媸媾嫁嫂嫉嫌嫒嫔嫖嫘嫜嫠嫡嫣嫦嫩嫫嫱嬉嬖嬗嬴嬷孀子孑孓孔孕字存孙孚孛孜孝孟孢季孤孥学孩孪孬孰孱孳孵孺孽宁它宄宅宇守安宋完宏宓宕宗官宙定宛宜宝实宠审客宣室宥宦宪宫宰害宴宵家宸容宽宾宿寂寄寅密寇富寐寒寓寝寞察寡寤寥寨寮寰寸对寺寻导寿封射将尉尊小少尔尕尖尘尚尜尝尤尥尧尬就尴尸尹尺尻尼尽尾尿局屁层居屈屉届屋屎屏屐屑展屙属屠屡屣履屦屯山屹屺屿岁岂岈岌岍岐岑岔岖岗岘岙岚岛岢岣岩岫岬岭岱岳岵岷岸岿峁峄峋峒峙峡峤峥峦峨峪峭峰峻崂崃崆崇崎崔崖崛崞崤崦崧崩崭崮崴崽嵇嵊嵋嵌嵘嵛嵝嵩嵫嵬嵯嵴嶂嶙嶝嶷巅巍川州巡巢工左巧巨巩巫差巯己已巳巴巷巽巾币市布帅帆师希帏帐帑帔帕帖帘帙帚帛帜帝带帧席帮帱帷常帻帼帽幂幄幅幌幔幕幛幞幡幢干平年并幸幺幻幼幽广庀庄庆庇床庋序庐庑库应底庖店庙庚府庞废庠庥度座庭庳庵庶康庸庹庾廉廊廒廓廖廛廨廪延廷建廿开弁异弃弄弈弊弋式弑弓引弗弘弛弟张弥弦弧弩弭弯弱弹强弼彀归当录彖彗彘彝形彤彦彩彪彬彭彰影彳彷役彻彼往征徂径待徇很徉徊律後徐徒徕得徘徙徜御徨循徭微徵德徼徽心必忆忉忌忍忏忐忑忒忖志忘忙忝忠忡忤忧忪快忭忮忱念忸忻忽忾忿怀态怂怃怄怅怆怊怍怎怏怒怔怕怖怙怛怜思怠怡急怦性怨怩怪怫怯怵总怼怿恁恂恃恋恍恐恒恕恙恚恝恢恣恤恧恨恩恪恫恬恭息恰恳恶恸恹恺恻恼恽恿悃悄悉悌悍悒悔悖悚悛悝悟悠患悦您悫悬悭悯悱悲悴悸悻悼情惆惊惋惑惕惘惚惜惝惟惠惦惧惨惩惫惬惭惮惯惰想惴惶惹惺愀愁愆愈愉愍愎意愕愚感愠愣愤愦愧愫愿慈慊慌慎慑慕慝慢慧慨慰慵慷憋憎憔憝憧憨憩憬憷憾懂懈懊懋懑懒懦懵懿戆戈戊戋戌戍戎戏成我戒戕或戗战戚戛戟戡戢戥截戬戮戳戴户戽戾房所扁扃扇扈扉手扌才扎扑扒打扔托扛扣扦执扩扪扫扬扭扮扯扰扳扶批扼找承技抄抉把抑抒抓投抖抗折抚抛抟抠抡抢护报抨披抬抱抵抹抻押抽抿拂拄担拆拇拈拉拊拌拍拎拐拒拓拔拖拗拘拙拚招拜拟拢拣拥拦拧拨择括拭拮拯拱拳拴拶拷拼拽拾拿持挂指挈按挎挑挖挚挛挝挞挟挠挡挢挣挤挥挨挪挫振挲挹挺挽捂捃捅捆捉捋捌捍捎捏捐捕捞损捡换捣捧捩捭据捱捶捷捺捻掀掂掇授掉掊掌掎掏掐排掖掘掠探掣接控推掩措掬掭掮掰掳掴掷掸掺掼掾揄揆揉揍揎描提插揖揞揠握揣揩揪揭援揶揸揽揿搀搁搂搅搋搌搏搐搓搔搛搜搞搠搡搦搪搬搭搴携搽摁摄摅摆摇摈摊摒摔摘摞摧摩摭摸摹摺撂撄撅撇撑撒撕撖撙撞撤撩撬播撮撰撵撷撸撺撼擀擂擅操擎擐擒擘擞擢擤擦攀攉攒攘攥攫攮支收攸改攻放政故效敉敌敏救敕敖教敛敝敞敢散敦敫敬数敲整敷文斋斌斐斑斓斗料斛斜斟斡斤斥斧斩斫断斯新方於施旁旃旄旅旆旋旌旎族旒旖旗无既日旦旧旨早旬旭旮旯旰旱时旷旺昀昂昃昆昊昌明昏易昔昕昙昝星映春昧昨昭是昱昴昵昶昼显晁晃晋晌晏晒晓晔晕晖晗晚晟晡晤晦晨普景晰晴晶晷智晾暂暄暇暌暑暖暗暝暧暨暮暴暹暾曙曛曜曝曦曩曰曲曳更曷曹曼曾替最月有朊朋服朐朔朕朗望朝期朦木未末本札术朱朴朵机朽杀杂权杆杈杉杌李杏材村杓杖杜杞束杠条来杨杪杭杯杰杲杳杵杷杼松板极构枇枉枋析枕林枘枚果枝枞枢枣枥枧枨枪枫枭枯枰枳枵架枷枸柁柃柄柏某柑柒染柔柘柙柚柜柝柞柠柢查柩柬柯柰柱柳柴柽柿栀栅标栈栉栊栋栌栎栏树栓栖栗栝校栩株栲栳样核根格栽栾桀桁桂桃桄桅框案桉桊桌桎桐桑桓桔桕桠桡桢档桤桥桦桧桨桩桫桴桶桷梁梃梅梆梏梓梗梢梦梧梨梭梯械梳梵检棂棉棋棍棒棕棘棚棠棣森棰棱棵棹棺棼椁椅椋植椎椐椒椟椠椤椭椰椴椹椽椿楂楔楗楚楝楞楠楣楦楫楮楷楸楹楼榀概榄榆榇榈榉榍榔榕榛榜榧榨榫榭榱榴榷榻槁槊槌槎槐槔槛槟槠槭槲槽槿樊樗樘樟模樨横樯樱樵樽樾橄橇橐橘橙橛橡橥橱橹橼檀檄檎檐檑檗檠檩檫檬欠次欢欣欤欧欲欷欺款歃歆歇歉歌歙止正此步武歧歪歹死歼殁殂殃殄殆殇殉殊残殍殒殓殖殚殛殡殪殳殴段殷殿毁毂毅毋母每毒毓比毕毖毗毙毛毡毪毫毯毳毵毹毽氅氆氇氍氏氐民氓气氕氖氘氙氚氛氟氡氢氤氦氧氨氩氪氮氯氰氲水永氽汀汁求汆汇汉汊汐汔汕汗汛汜汝汞江池污汤汨汩汪汰汲汴汶汹汽汾沁沂沃沅沆沈沉沌沏沐沓沔沙沛沟没沣沤沥沦沧沩沪沫沭沮沱河沸油治沼沽沾沿泄泅泉泊泌泐泓泔法泖泗泛泞泠泡波泣泥注泪泫泮泯泰泱泳泵泷泸泺泻泼泽泾洁洄洇洋洌洎洒洗洙洚洛洞津洧洪洫洮洱洲洳洵洹活洼洽派流浃浅浆浇浈浊测浍济浏浑浒浓浔浙浚浜浞浠浣浦浩浪浮浯浴海浸浼涂涅消涉涌涎涑涓涔涕涛涝涞涟涠涡涣涤润涧涨涩涪涫涮涯液涵涸涿淀淄淅淆淇淋淌淑淖淘淙淝淞淠淡淤淦淫淬淮深淳混淹添淼清渊渌渍渎渐渑渔渖渗渚渝渠渡渣渤渥温渫渭港渲渴游渺湃湄湍湎湔湖湘湛湟湫湮湾湿溃溅溆溉溏源溘溜溟溢溥溧溪溯溱溲溴溶溷溺溻溽滁滂滇滋滏滑滓滔滕滗滚滞滟滠满滢滤滥滦滨滩滴滹漂漆漉漏漓演漕漠漤漩漪漫漭漯漱漳漶漾潆潇潋潍潘潜潞潢潦潭潮潲潴潸潺潼澄澈澉澌澍澎澜澡澧澳澶澹激濂濉濑濒濞濠濡濮濯瀑瀚瀛瀣瀵瀹灌灏灞火灭灯灰灵灶灸灼灾灿炀炅炉炊炎炒炔炕炖炙炜炝炫炬炭炮炯炱炳炷炸点炻炼炽烀烁烂烃烈烊烘烙烛烟烤烦烧烨烩烫烬热烯烷烹烽焉焊焐焓焕焖焘焙焚焦焯焰焱然煅煊煌煎煜煞煤煦照煨煮煲煳煸煺煽熄熊熏熔熘熙熟熠熨熬熵熹燃燎燔燕燠燥燧燮燹爆爝爨爪爬爰爱爵父爷爸爹爻爽爿片版牌牍牒牖牙牛牝牟牡牢牦牧物牮牯牲牵特牺牾犀犁犄犊犋犍犏犒犟犬犯犰犴状犷犸犹狁狂狃狄狈狍狎狐狒狗狙狞狠狡狨狩独狭狮狯狰狱狲狳狴狷狸狺狻狼猁猃猊猎猕猖猗猛猜猝猞猡猢猥猩猪猫猬献猱猴猷猸猹猾猿獍獐獒獗獠獬獭獯獾玄率玉王玎玑玖玛玢玩玫玮环现玲玳玷玺玻珀珂珈珉珊珍珏珐珑珙珞珠珥珧珩班珲球琅理琉琏琐琚琛琢琥琦琨琪琬琮琰琳琴琵琶琼瑁瑕瑗瑙瑚瑛瑜瑞瑟瑭瑰瑶瑾璀璁璃璇璋璎璐璜璞璧璨璩瓒瓜瓞瓠瓢瓣瓤瓦瓮瓯瓴瓶瓷瓿甄甏甑甓甘甙甚甜生甥用甩甫甬甭田由甲申电男甸町画甾畀畅畈畋界畎畏畔留畚畛畜略畦番畲畴畸畹畿疃疆疋疏疑疔疖疗疙疚疝疟疠疡疣疤疥疫疬疮疯疰疱疲疳疴疵疸疹疼疽疾痂痃痄病症痈痉痊痍痒痔痕痘痛痞痢痣痤痦痧痨痪痫痰痱痴痹痼痿瘀瘁瘃瘅瘊瘌瘐瘗瘘瘙瘛瘟瘠瘢瘤瘥瘦瘩瘪瘫瘭瘰瘳瘴瘵瘸瘼瘾瘿癀癃癌癍癔癖癜癞癣癫癯癸登白百皂的皆皇皈皋皎皑皓皖皙皤皮皱皲皴皿盂盅盆盈益盍盎盏盐监盒盔盖盗盘盛盟盥目盯盱盲直相盹盼盾省眄眇眈眉看眍眙眚真眠眢眦眨眩眭眯眵眶眷眸眺眼着睁睃睇睐睑睚睛睡睢督睥睦睨睫睬睹睽睾睿瞀瞄瞅瞌瞍瞎瞑瞒瞟瞠瞢瞥瞧瞩瞪瞬瞰瞳瞵瞻瞽瞿矍矗矛矜矢矣知矧矩矫矬短矮石矶矸矽矾矿砀码砂砉砌砍砑砒研砖砗砘砚砜砝砟砣砥砧砭砰破砷砸砹砺砻砼砾础硅硇硌硎硐硒硕硖硗硝硪硫硬硭确硷硼碇碉碌碍碎碑碓碗碘碚碛碜碟碡碣碥碧碰碱碲碳碴碹碾磁磅磉磊磋磐磔磕磙磨磬磲磴磷磺礁礅礓礞礤礴示礻礼社祀祁祆祈祉祓祖祗祚祛祜祝神祟祠祢祥祧票祭祯祷祸祺禀禁禄禅禊福禚禧禳禹禺离禽禾秀私秃秆秉秋种科秒秕秘租秣秤秦秧秩秫秭积称秸移秽稀稂稆程稍税稔稗稚稞稠稣稳稷稻稼稽稿穆穑穗穰穴究穷穸穹空穿窀突窃窄窈窍窑窒窕窖窗窘窜窝窟窠窥窦窨窬窭窳窿立竖站竞竟章竣童竦竭端竹竺竽竿笃笄笆笈笊笋笏笑笔笕笙笛笞笠笤笥符笨笪笫第笮笱笳笸笺笼笾筅筇等筋筌筏筐筑筒答策筘筚筛筝筠筢筮筱筲筵筷筹筻签简箅箍箐箔箕算箜管箢箦箧箨箩箪箫箬箭箱箴箸篁篆篇篌篑篓篙篚篝篡篥篦篪篮篱篷篼篾簇簋簌簏簖簟簦簧簪簸簿籀籁籍米籴类籼籽粉粑粒粕粗粘粜粝粞粟粤粥粪粮粱粲粳粹粼粽精糁糅糇糈糊糌糍糕糖糗糙糜糟糠糨糯系紊素索紧紫累絮絷綦綮縻繁繇纂纛纠纡红纣纤纥约级纨纩纪纫纬纭纯纰纱纲纳纵纶纷纸纹纺纽纾线绀绁绂练组绅细织终绉绊绋绌绍绎经绐绑绒结绔绕绗绘给绚绛络绝绞统绠绡绢绣绥绦继绨绩绪绫续绮绯绰绲绳维绵绶绷绸绺绻综绽绾绿缀缁缂缃缄缅缆缇缈缉缌缎缏缑缒缓缔缕编缗缘缙缚缛缜缝缟缠缡缢缣缤缥缦缧缨缩缪缫缬缭缮缯缰缱缲缳缴缵缶缸缺罂罄罅罐网罔罕罗罘罚罟罡罢罨罩罪置罱署罴罹罾羁羊羌美羔羚羝羞羟羡群羧羯羰羲羸羹羼羽羿翁翅翊翌翎翔翕翘翟翠翡翥翦翩翮翰翱翳翻翼耀老考耄者耆耋而耍耐耒耔耕耖耗耘耙耜耠耢耥耦耧耨耩耪耱耳耵耶耷耸耻耽耿聂聃聆聊聋职聍聒联聘聚聩聪聱聿肃肄肆肇肉肋肌肓肖肘肚肛肝肟肠股肢肤肥肩肪肫肭肮肯肱育肴肷肺肼肽肾肿胀胁胂胃胄胆背胍胎胖胗胙胚胛胜胝胞胡胤胥胧胨胩胪胫胬胭胯胰胱胲胳胴胶胸胺胼能脂脆脉脊脍脎脏脐脑脒脓脔脖脘脚脞脬脯脱脲脶脸脾腆腈腊腋腌腐腑腓腔腕腙腚腠腥腧腩腭腮腰腱腴腹腺腻腼腽腾腿膀膂膈膊膏膑膘膛膜膝膦膨膪膳膺膻臀臁臂臃臆臊臌臣臧自臬臭至致臻臼臾舀舁舂舄舅舆舌舍舐舒舔舛舜舞舟舡舢舣舨航舫般舰舱舳舴舵舶舷舸船舻舾艄艇艋艘艚艟艨艮良艰色艳艴艺艽艾艿节芄芈芊芋芍芎芏芑芒芗芘芙芜芝芟芡芥芦芨芩芪芫芬芭芮芯芰花芳芴芷芸芹芽芾苁苄苇苈苊苋苌苍苎苏苑苒苓苔苕苗苘苛苜苞苟苠苡苣苤若苦苫苯英苴苷苹苻茁茂范茄茅茆茈茉茌茎茏茑茔茕茗茚茛茜茧茨茫茬茭茯茱茳茴茵茶茸茹茼荀荃荆荇草荏荐荑荒荔荚荛荜荞荟荠荡荣荤荥荦荧荨荩荪荫荬荭药荷荸荻荼荽莅莆莉莎莒莓莘莛莜莞莠莨莩莪莫莰莱莲莳莴莶获莸莹莺莼莽菀菁菅菇菊菌菏菔菖菘菜菝菟菠菡菥菩菪菰菱菲菹菽萁萃萄萋萌萍萎萏萑萘萜萝萤营萦萧萨萱萸萼落葆葑著葚葛葜葡董葩葫葬葭葱葳葵葶葸葺蒂蒇蒈蒉蒋蒌蒎蒗蒙蒜蒡蒯蒲蒴蒸蒹蒺蒽蒿蓁蓄蓉蓊蓍蓐蓑蓓蓖蓝蓟蓠蓣蓥蓦蓬蓰蓼蓿蔌蔑蔓蔗蔚蔟蔡蔫蔬蔷蔸蔹蔺蔻蔼蔽蕃蕈蕉蕊蕖蕙蕞蕤蕨蕲蕴蕹蕺蕻蕾薄薅薇薏薛薜薤薨薪薮薯薰薷薹藁藉藏藐藓藕藜藤藩藻藿蘅蘑蘖蘧蘩蘸蘼虎虏虐虑虔虚虞虢虫虬虮虱虹虺虻虼虽虾虿蚀蚁蚂蚊蚋蚌蚍蚓蚕蚜蚝蚣蚤蚧蚨蚩蚬蚯蚰蚱蚴蚶蚺蛀蛄蛆蛇蛉蛊蛋蛎蛏蛐蛑蛔蛘蛙蛛蛞蛟蛤蛩蛭蛮蛰蛱蛲蛳蛴蛸蛹蛾蜀蜂蜃蜇蜈蜉蜊蜍蜒蜓蜕蜗蜘蜚蜜蜞蜡蜢蜣蜥蜩蜮蜱蜴蜷蜻蜾蜿蝇蝈蝉蝌蝎蝓蝗蝙蝠蝣蝤蝥蝮蝰蝴蝶蝻蝼蝽蝾螂螃螅螈螋融螗螟螨螫螬螭螯螳螵螺螽蟀蟆蟊蟋蟑蟒蟛蟠蟥蟪蟮蟹蟾蠃蠊蠓蠕蠖蠡蠢蠲蠹蠼血衄衅行衍衔街衙衡衢衣补表衩衫衬衮衰衲衷衽衾衿袁袂袄袅袈袋袍袒袖袜袢袤被袭袱袼裁裂装裆裉裎裒裔裕裘裙裟裢裣裤裥裨裰裱裳裴裸裹裼裾褂褊褐褒褓褙褚褛褡褥褪褫褰褴褶襁襄襞襟襦襻西要覃覆见观规觅视觇览觉觊觋觌觎觏觐觑角觖觚觜觞解觥触觫觯觳言訇訾詈詹誉誊誓謇警譬计订讣认讥讦讧讨让讪讫训议讯记讲讳讴讵讶讷许讹论讼讽设访诀证诂诃评诅识诈诉诊诋诌词诎诏译诒诓诔试诖诗诘诙诚诛诜话诞诟诠诡询诣诤该详诧诨诩诫诬语诮误诰诱诲诳说诵请诸诹诺读诼诽课诿谀谁谂调谄谅谆谇谈谊谋谌谍谎谏谐谑谒谓谔谕谖谗谙谚谛谜谝谟谠谡谢谣谤谥谦谧谨谩谪谫谬谭谮谯谰谱谲谳谴谵谶谷豁豆豇豉豌豕豚象豢豪豫豳豸豹豺貂貅貉貊貌貔貘贝贞负贡财责贤败账货质贩贪贫贬购贮贯贰贱贲贳贴贵贶贷贸费贺贻贼贽贾贿赀赁赂赃资赅赆赇赈赉赊赋赌赍赎赏赐赓赔赖赘赙赚赛赜赝赞赠赡赢赣赤赦赧赫赭走赳赴赵赶起趁趄超越趋趑趔趟趣趱足趴趵趸趺趼趾趿跃跄跆跋跌跎跏跑跖跗跚跛距跞跟跣跤跨跪跬路跳践跷跸跹跺跻跽踅踉踊踌踏踔踝踞踟踢踣踩踪踬踮踯踱踵踹踺踽蹀蹁蹂蹄蹇蹈蹉蹊蹋蹑蹒蹙蹦蹩蹬蹭蹯蹰蹲蹴蹶蹼蹿躁躅躇躏躐躔躜躞身躬躯躲躺车轧轨轩轫转轭轮软轰轱轲轳轴轵轶轷轸轺轻轼载轾轿辁辂较辄辅辆辇辈辉辊辋辍辎辏辐辑输辔辕辖辗辘辙辚辛辜辞辟辣辨辩辫辰辱边辽达迁迂迄迅过迈迎运近迓返迕还这进远违连迟迢迤迥迦迨迩迪迫迭迮述迷迸迹追退送适逃逄逅逆选逊逋逍透逐逑递途逖逗通逛逝逞速造逡逢逦逭逮逯逵逶逸逻逼逾遁遂遄遇遍遏遐遑遒道遗遘遛遢遣遥遨遭遮遴遵遽避邀邂邃邈邋邑邓邕邗邙邛邝邡邢那邦邪邬邮邯邰邱邳邴邵邶邸邹邺邻邾郁郄郅郇郊郎郏郐郑郓郗郛郜郝郡郢郦郧部郫郭郯郴郸都郾鄂鄄鄙鄞鄢鄣鄯鄱鄹酃酆酉酊酋酌配酎酏酐酒酗酚酝酞酡酢酣酤酥酩酪酬酮酯酰酱酲酴酵酶酷酸酹酽酾酿醅醇醉醋醌醍醐醑醒醚醛醢醪醭醮醯醴醵醺采釉释里重野量金釜鉴銎銮鋈錾鍪鎏鏊鏖鐾鑫钆钇针钉钊钋钌钍钎钏钐钒钓钔钕钗钙钚钛钜钝钞钟钠钡钢钣钤钥钦钧钨钩钪钫钬钭钮钯钰钱钲钳钴钵钷钹钺钻钼钽钾钿铀铁铂铃铄铅铆铈铉铊铋铌铍铎铐铑铒铕铗铘铙铛铜铝铞铟铠铡铢铣铤铥铧铨铩铪铫铬铭铮铯铰铱铲铳铴铵银铷铸铹铺铼铽链铿销锁锂锃锄锅锆锇锈锉锊锋锌锎锏锐锑锒锓锔锕锖锗锘错锚锛锝锞锟锡锢锣锤锥锦锨锩锪锫锬锭键锯锰锱锲锴锵锶锷锸锹锺锻锾锿镀镁镂镄镅镆镇镉镊镌镍镎镏镐镑镒镓镔镖镗镘镛镜镝镞镡镢镣镤镥镦镧镨镩镪镫镬镭镯镰镱镲镳镶长门闩闪闫闭问闯闰闱闲闳间闵闶闷闸闹闺闻闼闽闾阀阁阂阃阄阅阆阈阉阊阋阌阍阎阏阐阑阒阔阕阖阗阙阚阜队阡阢阪阮阱防阳阴阵阶阻阼阽阿陀陂附际陆陇陈陉陋陌降限陔陕陛陟陡院除陧陨险陪陬陲陴陵陶陷隅隆隈隋隍随隐隔隗隘隙障隧隰隳隶隼隽难雀雁雄雅集雇雉雌雍雎雏雒雕雠雨雩雪雯雳零雷雹雾需霁霄霆震霈霉霍霎霏霓霖霜霞霪霭霰露霸霹霾青靓靖静靛非靠靡面靥革靳靴靶靼鞅鞋鞍鞑鞒鞘鞠鞣鞫鞭鞯鞲鞴韦韧韩韪韫韬韭音韵韶页顶顷顸项顺须顼顽顾顿颀颁颂颃预颅领颇颈颉颊颌颍颏颐频颓颔颖颗题颚颛颜额颞颟颠颡颢颤颥颦颧风飑飒飓飕飘飙飚飞食飧飨餍餐餮饔饕饥饧饨饩饪饫饬饭饮饯饰饱饲饴饵饶饷饺饼饽饿馀馁馄馅馆馇馈馊馋馍馏馐馑馒馓馔馕首馗馘香馥馨马驭驮驯驰驱驳驴驵驶驷驸驹驺驻驼驽驾驿骀骁骂骄骅骆骇骈骊骋验骏骐骑骒骓骖骗骘骚骛骜骝骞骟骠骡骢骣骤骥骧骨骰骶骷骸骺骼髀髁髂髅髋髌髑髓高髡髦髫髭髯髹髻鬃鬈鬏鬓鬟鬣鬯鬲鬻鬼魁魂魃魄魅魇魈魉魍魏魑魔鱼鱿鲁鲂鲅鲆鲇鲈鲋鲍鲎鲐鲑鲔鲚鲛鲜鲞鲟鲠鲡鲢鲣鲤鲥鲦鲧鲨鲩鲫鲭鲮鲰鲱鲲鲳鲴鲵鲷鲸鲺鲻鲼鲽鳃鳄鳅鳆鳇鳊鳌鳍鳎鳏鳐鳓鳔鳕鳖鳗鳘鳙鳜鳝鳞鳟鳢鸟鸠鸡鸢鸣鸥鸦鸨鸩鸪鸫鸬鸭鸯鸱鸲鸳鸵鸶鸷鸸鸹鸺鸽鸾鸿鹁鹂鹃鹄鹅鹆鹇鹈鹉鹊鹋鹌鹎鹏鹑鹕鹗鹘鹚鹛鹜鹞鹣鹤鹦鹧鹨鹩鹪鹫鹬鹭鹰鹱鹳鹿麂麇麈麋麒麓麝麟麦麸麻麽麾黄黉黍黎黏黑黔默黛黜黝黟黠黢黥黧黩黯黹黻黼黾鼋鼍鼎鼐鼓鼗鼙鼠鼢鼬鼯鼷鼹鼻鼾齐齑齿龀龃龄龅龆龇龈龉龊龋龌龙龚龛龟龠',
+  korean:
+    ' !"#$%&\'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\\]^_`abcdefghijklmnopqrstuvwxyz{|}~가각간갇갈감갑값강갖같갚갛개객걀거걱건걷걸검겁것겉게겨격겪견결겹경곁계고곡곤곧골곰곱곳공과관광괜괴굉교구국군굳굴굵굶굽궁권귀규균그극근글긁금급긋긍기긴길김깅깊까깎깐깔깜깝깥깨꺼꺾껍껏껑께껴꼬꼭꼴꼼꼽꽂꽃꽉꽤꾸꿀꿈뀌끄끈끊끌끓끔끗끝끼낌나낙낚난날낡남납낫낭낮낯낱낳내냄냉냐냥너넉널넓넘넣네넥넷녀녁년념녕노녹논놀놈농높놓놔뇌뇨누눈눕뉘뉴늄느늑는늘늙능늦늬니닐님다닥닦단닫달닭닮담답닷당닿대댁댐더덕던덜덤덥덧덩덮데델도독돈돌돕동돼되된두둑둘둠둡둥뒤뒷드득든듣들듬듭듯등디딩딪따딱딴딸땀땅때땜떠떡떤떨떻떼또똑뚜뚫뚱뛰뜨뜩뜯뜰뜻띄라락란람랍랑랗래랜램랫략량러럭런럴럼럽럿렁렇레렉렌려력련렬렵령례로록론롬롭롯료루룩룹룻뤄류륙률륭르른름릇릎리릭린림립릿마막만많말맑맘맙맛망맞맡맣매맥맨맵맺머먹먼멀멈멋멍멎메멘멩며면멸명몇모목몰몸몹못몽묘무묵묶문묻물뭄뭇뭐뭣므미민믿밀밉밌및밑바박밖반받발밝밟밤밥방밭배백뱀뱃뱉버번벌범법벗베벤벼벽변별볍병볕보복볶본볼봄봇봉뵈뵙부북분불붉붐붓붕붙뷰브블비빌빗빚빛빠빨빵빼뺨뻐뻔뻗뼈뽑뿌뿐쁘쁨사삭산살삶삼상새색샌생서석섞선설섬섭섯성세센셈셋션소속손솔솜솟송솥쇄쇠쇼수숙순술숨숫숲쉬쉽슈스슨슬슴습슷승시식신싣실싫심십싱싶싸싹쌀쌍쌓써썩썰썹쎄쏘쏟쑤쓰쓸씀씌씨씩씬씹씻아악안앉않알앓암압앗앙앞애액야약얇양얗얘어억언얹얻얼엄업없엇엉엌엎에엔엘여역연열엷염엽엿영옆예옛오옥온올옮옳옷와완왕왜왠외왼요욕용우욱운울움웃웅워원월웨웬위윗유육율으윽은을음응의이익인일읽잃임입잇있잊잎자작잔잖잘잠잡장잦재쟁저적전절젊점접젓정젖제젠젯져조족존졸좀좁종좋좌죄주죽준줄줌줍중쥐즈즉즌즐즘증지직진질짐집짓징짙짚짜짝짧째쨌쩌쩍쩐쪽쫓쭈쭉찌찍찢차착찬찮찰참창찾채책챔챙처척천철첫청체쳐초촉촌총촬최추축춘출춤춥춧충취츠측츰층치칙친칠침칭카칸칼캐캠커컨컬컴컵컷켓켜코콜콤콩쾌쿠퀴크큰클큼키킬타탁탄탈탑탓탕태택탤터턱털텅테텍텔템토톤톱통퇴투툼퉁튀튜트특튼튿틀틈티틱팀팅파팎판팔패팩팬퍼퍽페펴편펼평폐포폭표푸푹풀품풍퓨프플픔피픽필핏핑하학한할함합항해핵핸햄햇행향허헌험헤헬혀현혈협형혜호혹혼홀홍화확환활황회획횟효후훈훌훔훨휘휴흉흐흑흔흘흙흡흥흩희흰히힘',
+  telugu:
+    '0123456789!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~ abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZఁంఃఅఆఇఈఉఊఋఌఎఏఐఒఓఔకఖగఘఙచఛజఝఞటఠడఢణతథదధనపఫబభమయరఱలళవశషసహాిీుూృౄెేైొోౌ్ౠౡౢౣ',
+  kannada:
+    '0123456789!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~ abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZಂಃಅಆಇಈಉಊಋಎಏಐಒಓಔಕಖಗಘಙಚಛಜಝಞಟಠಡಢಣತಥದಧನಪಫಬಭಮಯರಲಳವಶಷಸಹಾಿೀುೂೃೆೇೈೊೋೌ್೦೧೨೩೪೫೬೭೮೯',
+};
+
+/**
+ * Mapping of language codes to their corresponding symbol sets.
+ */
+export const symbols = {
+  // Abaza
+  abq: alphabets.cyrillic,
+  // Adyghe
+  ady: alphabets.cyrillic,
+  // Africans
+  af: alphabets.latin,
+  // Avar
+  ava: alphabets.cyrillic,
+  // Azerbaijani
+  az: alphabets.latin,
+  // Belarusian
+  be: alphabets.cyrillic,
+  // Bulgarian
+  bg: alphabets.cyrillic,
+  // Bosnian
+  bs: alphabets.latin,
+  // Simplified Chinese
+  chSim: alphabets.zh_sim,
+  // Chechen
+  che: alphabets.cyrillic,
+  // Czech
+  cs: alphabets.latin,
+  // Welsh
+  cy: alphabets.latin,
+  // Danish
+  da: alphabets.latin,
+  // Dargwa
+  dar: alphabets.cyrillic,
+  // German
+  de: alphabets.latin,
+  // English
+  en: alphabets.english,
+  // Spanish
+  es: alphabets.latin,
+  // Estonian
+  et: alphabets.latin,
+  // French
+  fr: alphabets.latin,
+  // Irish
+  ga: alphabets.latin,
+  // Croatian
+  hr: alphabets.latin,
+  // Hungarian
+  hu: alphabets.latin,
+  // Indonesian
+  id: alphabets.latin,
+  // Ingush
+  inh: alphabets.cyrillic,
+  // Icelandic
+  ic: alphabets.latin,
+  // Italian
+  it: alphabets.latin,
+  // Japanese
+  ja: alphabets.japanese,
+  // Karbadian
+  kbd: alphabets.cyrillic,
+  // Kannada
+  kn: alphabets.kannada,
+  // Korean
+  ko: alphabets.korean,
+  // Kurdish
+  ku: alphabets.latin,
+  // Latin
+  la: alphabets.latin,
+  // Lak
+  lbe: alphabets.cyrillic,
+  // Lezghian
+  lez: alphabets.cyrillic,
+  // Lithuanian
+  lt: alphabets.latin,
+  // Latvian
+  lv: alphabets.latin,
+  // Maori
+  mi: alphabets.latin,
+  // Mongolian
+  mn: alphabets.cyrillic,
+  // Malay
+  ms: alphabets.latin,
+  // Maltese
+  mt: alphabets.latin,
+  // Dutch
+  nl: alphabets.latin,
+  // Norwegian
+  no: alphabets.latin,
+  // Occitan
+  oc: alphabets.latin,
+  // Pali
+  pi: alphabets.latin,
+  // Polish
+  pl: alphabets.latin,
+  // Portuguese
+  pt: alphabets.latin,
+  // Romanian
+  ro: alphabets.latin,
+  // Russian
+  ru: alphabets.cyrillic,
+  // Serbian (cyrillic)
+  rsCyrillic: alphabets.cyrillic,
+  // Serbian (latin)
+  rsLatin: alphabets.latin,
+  // Slovak
+  sk: alphabets.latin,
+  // Slovenian
+  sl: alphabets.latin,
+  // Albanian
+  sq: alphabets.latin,
+  // Swedish
+  sv: alphabets.latin,
+  // Swahili
+  sw: alphabets.latin,
+  // Tabassaran
+  tab: alphabets.cyrillic,
+  // Telugu
+  te: alphabets.telugu,
+  // Tajik
+  tjk: alphabets.cyrillic,
+  // Tagalog
+  tl: alphabets.latin,
+  // Turkish
+  tr: alphabets.latin,
+  // Ukrainian
+  uk: alphabets.cyrillic,
+  // Uzbek
+  uz: alphabets.latin,
+  // Vietnamese
+  vi: alphabets.latin,
+};
+
+/**
+ * Supported OCR language codes (EasyOCR alphabets).
+ * @category Types
+ */
+export type OCRLanguage = keyof typeof symbols;
+
+export const PPOCR_SYMBOLS =
+  '!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz¢£¤¥¦§¨©ª«¬®¯°±²³´µ¶·¸¹º»¼½¾¿ÀÁÂÃÄÅÆÇÈÉÊËÌÍÎÏÐÑÒÓÔÕÖ×ØÙÚÛÜÝÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüýþÿĀāĂ㥹ĆćĈĉĊċČčĎďĐđĒēĔĕĖėĘęĚěĜĝĞğĠġĢģĤĥĦħĨĩĪīĬĭĮįİıIJijĴĵĶķĸĹĺĻļĽľĿŀŁłŃńŅņŇňʼnŊŋŌōŎŏŐőŒœŔŕŖŗŘřŚśŜŝŞşŠšŢţŤťŦŧŨũŪūŬŭŮůŰűŲųŴŵŶŷŸŹźŻżŽžſƀƁƂƃƄƅƆƇƈƉƊƋƌƍƎƏƐƑƒƓƔƕƖƗƘƙƚƛƜƝƞƟƠơƢƣƤƥƦƧƨƩƪƫƬƭƮƯưƱƲƳƴƵƶƷƸƹƺƻƼƽƾƿǀǁǂǃDŽDždžLJLjljNJNjnjǍǎǏǐǑǒǓǔǕǖǗǘǙǚǛǜǝǞǟǠǡǢǣǤǥǦǧǨǩǪǫǬǭǮǯǰDZDzdzǴǵǶǷǸǹǺǻǼǽǾǿȀȁȂȃȄȅȆȇȈȉȊȋȌȍȎȏȐȑȒȓȔȕȖȗȘșȚțȜȝȞȟȠȡȢȣȤȥȦȧȨȩȪȫȬȭȮȯȰȱȲȳȴȵȶȷȸȹȺȻȼȽȾȿɀɁɂɃɄɅɆɇɈɉɊɋɌɍɎɏɐɑɒɓɔɕɖɗɘəɚɛɜɝɞɟɠɡɢɣɤɥɦɧɨɩɪɫɬɭɮɯɰɱɲɳɴɵɶɷɸɹɺɻɼɽɾɿʀʁʂʃʄʅʆʇʈʉʊʋʌʍʎʏʐʑʒʓʔʕʖʗʘʙʚʛʜʝʞʟʠʡʢʣʤʥʦʧʨʩʪʫʬʭʮʯΑΒΓΔΕΖΗΘΙΚΛΜΝΞΟΠΡΣΤΥΦΧΨΩαβγδεζηθικλμνξοπρστυφχψωϐϑϒϕϖϝϞϟϠϡϢϣϤϥϦϧϨϩϪϫϬϭϮϯϰϱϴϵẠồ–—―‖‘’“”†‡•‥…‰′″※⁎⁰⁴⁵⁶⁷⁸⁹⁺⁻⁼⁽⁾ⁿ₀₁₂₃₄₅₆₇₈₉₊₋₌₍₎ₒ₠₡₢₣₤₥₦₧₨₩₪₫€₭₮₯₰₱₲₳₴₵₶₷₸₹₺₻₼₽₾₿℃℉ℎℏℑ℘ℜ™℧Åℵℶℷℸℹ⅀ⅠⅡⅢⅣⅤⅥⅦⅧⅨⅩⅪⅫⅰⅱⅲⅳⅴⅵⅶⅷⅸⅹⅺⅻⅼↀↁↂ←↑→↓↔↕↖↗↘↙↚↛↜↝↞↟↠↡↢↣↤↥↦↧↨↩↪↫↬↭↮↯↰↱↲↳↴↵↶↷↸↹↺↻↼↽↾↿⇀⇁⇂⇃⇄⇅⇆⇇⇈⇉⇊⇋⇌⇍⇎⇏⇐⇑⇒⇓⇔⇕⇖⇗⇘⇙⇚⇛⇜⇝⇞⇟⇠⇡⇢⇣⇤⇥⇦⇧⇨⇩⇪⇫⇬⇭⇮⇯⇰⇱⇲⇳⇴⇵⇶⇷⇸⇹⇺⇻⇼⇽⇾⇿∀∁∂∃∄∅∆∇∈∉∋∏∑−∓∕∖∙√∛∜∝∞∟∠∡∢∥∧∨∩∪∫∬∭∮∯∰∱∲∳∴∵∶∷∸∹∺∻∼∽∾∿≀≁≂≃≄≅≆≇≈≉≊≋≌≍≎≏≐≑≒≓≔≕≖≗≘≙≚≛≜≝≞≟≠≡≢≣≤≥≦≧≨≩≪≫≬≭≮≯≰≱≲≳≴≵≶≷≸≹≺≻≼≽≾≿⊀⊁⊂⊃⊄⊅⊆⊇⊈⊉⊊⊋⊌⊍⊎⊏⊐⊑⊒⊓⊔⊕⊖⊗⊘⊙⊚⊛⊜⊝⊞⊟⊠⊡⊢⊣⊤⊥⊦⊧⊨⊩⊪⊫⊬⊭⊮⊯⊰⊱⊲⊳⊴⊵⊶⊷⊸⊹⊺⊻⊼⊽⊾⊿⋅⌀⌃⌘⌚⌛⌤⌥⌦⌧⌨〉⌫⌬⌭⌮⌯⍵⍺⎆⎇⎈⎉⎊⎋⎌⎍⎎⎏⎐⎑⎒⎓⎔⎕⎖⎗⎘⎙⎚⎛⎜⎝⎞⎟⎠⎡⎢⎣⎤⎥⎦⎧⎨⎩⎪⎫⎬⎭⎮⎯⎰⎱⎲⎳⎴⎵⎶⎷⎸⎹⎺⎻⎼⎽⎾⎿⏀⏁⏂⏃⏄⏅⏆⏇⏈⏉⏊⏋⏌⏍⏎⏏⏐⏑⏒⏓⏔⏕⏖⏗⏘⏙⏚⏛⏜⏝⏞⏟⏠⏡⏢⏣⏤⏥⏦⏧⏨⏩⏪⏫⏬⏭⏮⏯⏰⏱⏲⏳⏴⏵⏶⏷⏸⏹⏺⏻⏼⏽⏾⏿①②③④⑤⑥⑦⑧⑨⑩⑪⑫⑬⑭⑮⑯⑰⑱⑲⑳Ⓜ━╆╳▁█■□▪▫▬▲△▶▷▼▽◀◆◇◉◊○◎●◐◥◻◼◽◾☀☁☂☃☄★☆☇☈☉☊☋☌☍☎☏☐☑☒☓☔☕☖☗☘☙☚☛☜☝☞☟☠☡☢☣☤☥☦☧☨☩☪☫☬☭☮☯☰☱☲☳☴☵☶☷☸☹☺☻☼☽☾☿♀♁♂♃♄♅♆♇♈♉♊♋♌♍♎♏♐♑♒♓♔♕♖♗♘♙♚♛♜♝♞♟♠♡♢♣♤♥♦♧♨♩♪♫♬♭♮♯♰♱♲♳♴♵♶♷♸♹♺♻♼♽♾♿⚀⚁⚂⚃⚄⚅⚆⚇⚈⚉⚊⚋⚌⚍⚎⚏⚐⚑⚒⚓⚔⚕⚖⚗⚘⚙⚚⚛⚜⚝⚞⚟⚠⚡⚢⚣⚤⚥⚦⚧⚨⚩⚪⚫⚬⚭⚮⚯⚰⚱⚲⚳⚴⚵⚶⚷⚸⚹⚺⚻⚼⚽⚾⚿⛀⛁⛂⛃⛄⛅⛆⛇⛈⛉⛊⛋⛌⛍⛎⛏⛐⛑⛒⛓⛔⛕⛖⛗⛘⛙⛚⛛⛜⛝⛞⛠⛡⛢⛣⛤⛥⛦⛧⛨⛩⛪⛫⛬⛭⛮⛯⛲⛳⛵⛶⛺⛽⛾⛿✂✅✆✇✈✉✊✋✌✍✎✏✐✑✒✓✔✕✖✘✙✚✛✜✝✞✟✠✡✢✣✤✥✦✧✨✩✪✫✬✭✮✯✰✱✲✳✴✵✶✷✸✹✺✻✼✽✾✿❀❁❂❃❄❅❆❇❈❉❊❋❌❍❎❏❐❑❒❓❔❕❖❗❘❙❚❛❜❝❞❡❢❣❤❥❦❧❨❩❪❫❬❭❮❯❰❱❲❳❴❵❶❷❸❹❺❻❼❽❾❿➀➁➂➃➄➅➆➇➈➉➊➋➌➍➎➏➐➑➒➓➔➕➖➗➘➙➚➛➜➝➞➟➠➡➢➣➤➥➦➧➨➩➪➫➬➭➮➯➰➱➲➳➴➵➶➷➸➹➺➻➼➽➾➿⟶⟹⤴⤵⬅⬆⬇⬛⬜⭐⭕　、。々〆〇〈〉《》「」『』【】〒〔〕〖〗〜〰〵〽ぁあぃいぅうぇえぉおかがきぎくぐけげこごさざしじすずせぜそぞただちぢっつづてでとどなにぬねのはばぱひびぴふぶぷへべぺほぼぽまみむめもゃやゅゆょよらりるれろゎわゑをんゕゖゝゞァアィイゥウェエォオカガキギクグケゲコゴサザシジスズセゼソゾタダチヂッツヅテデトドナニヌネノハバパヒビピフブプヘベペホボポマミムメモャヤュユョヨラリルレロヮワヰヲンヴヵヶヷヸヹヺ・ーヽヾヿ㈱㊗㊙㐂㐱㑇㑊㑳㑺㒺㕁㕑㕥㕮㘎㘭㙍㙘㙟㙦㚷㛃㛚㛹㝛㝠㝡㟁㟃㠇㠓㠣㠯㠶㡌㢘㤘㤙㥄㥫㥮㧐㧑㧟㧱㨗㨪㩗㩦㩳㪚㪟㫰㬉㬊㬎㬚㬪㭎㭕㮣㮾㰀㳄㳇㳒㳘㳚㴔㴪㴱㵐㶲㸃㸆㸌㺄㻬㼝㽏㽞㿠䁖䂮䃅䃎䃮䅟䈰䊀䌹䎃䎖䏝䏡䏲䐃䓖䓛䓣䓨䓫䓬䖝䗖䗛䗪䗬䗴䘏䘑䘚䜣䝉䝔䝙䠀䠶䢺䢼䣘䥽䦃䮄䰟䰾䲁䲟䲠䲢䴉䴓䴔䴕䴖䴗䴘䴙䶊䶮一丁七丄万丈三上下丌不与丏丐丑专且丕世丘丙业丛东丝丞丟両丢两严並丧个丫中丮丰丱串临丸丹为主丼丽举丿乂乃久么义之乌乍乎乏乐乒乓乔乖乗乘乙乜九乞也习乡书乩乭买乱乳乸乹乾亀亂了予争亊事二亍于亏云互亓五井亘亙亚些亜亞亟亡亢交亥亦产亨亩享京亭亮亯亰亱亲亳亵亶亷亸亹人亾亿什仁仂仃仄仅仆仇仉今介仍从仏仑仓仔仕他仗付仙仛仜仝仞仟仡代令以仨仩仪仫们仮仰仱仲仳仵件价仺任仼份仿伀企伃伈伉伊伋伍伎伏伐休伔伕众优伙会伛伝伞伟传伢伣伤伥伦伧伪伫伬伭伯估伱伲伴伶伷伸伺伻似伽伾佀佁佃但佇佈佉佋佌位低住佐佑佒体佔何佖佗佘余佚佛作佝佞佟你佢佣佤佥佧佩佪佬佯佰佳佴併佶佷佸佹佺佻佼佽佾使侀侁侂侃侄侅來侇侈侉侊例侍侏侐侑侔侖侗侘侚供侜依侞侠価侣侥侦侧侨侩侪侬侮侯侲侳侴侵侶侷侹侻便俁係促俄俅俇俉俊俋俌俍俎俏俐俑俓俔俗俘俙俚俛俜保俞俟俠信俣俤俦俨俩俪俫俬俭修俯俱俳俴俵俶俷俸俺俽俾倀倃倅倆倇倈倉個倌倍倏倐們倒倓倔倕倖倗倘候倚倛倜倞借倠倡倢倣値倥倦倧倨倩倪倫倬倭倮倰倳倴倶倷倸倹债倻值倾偁偃偅假偈偉偊偋偌偍偎偏偐偓偕偘做偛停偝偞偟偠偡偢健偨偩偪偫偬偭偯偰偲側偵偶偷偸偺偻偽偾偿傀傂傃傅傈傉傋傌傍傎傑傒傔傕傖傘備傚傛傜傝傞傢傣傥傧储傩催傭傮傯傰傱傲傳傴債傷傺傻傽傾傿僁僂僄僅僆僇僈僉僊僋働僎像僑僓僔僕僖僗僙僚僛僜僝僞僡僤僥僦僧僨僩僪僬僭僮僰僱僳僴僵僶僸價僻僽僾僿儀儁儂儃億儆儇儈儉儋儌儐儑儒儓儔儕儗儘儚儛儜儞償儠儡儢儤儥儦優儫儰儱儲儳儴儵儷儸儹儺儻儼儽儿兀允元兄充兆兇先光克兌免兎児兑兒兔兕兖兗兙党兛兜兝兞兠兡兢兣入內全兩兪八公六兮兰共关兴兵其具典兹养兼兽兿冀冁冂冄内円冇冈冉冊册再冏冐冑冒冓冔冕冗冘写冚军农冞冠冢冣冤冥冧冨冪冬冮冯冰冱冲决冴况冶冷冺冻冼冽净凃凄凅准凇凈凉凊凋凌凍减凑凔凖凘凛凜凝凞几凡凢凤処凪凫凬凭凯凰凱凳凵凶凸凹出击凼函凿刀刁刂刃刄分切刈刉刊刌刍刎刑划刓刖列刘则刚创刜初刞删判別刦刧刨利刪别刬刭刮到刱刲刳刵制刷券刹刺刻刼刽刿剀剁剂剃剄剅剆則剉削剋剌前剎剏剐剑剒剔剕剖剗剚剛剜剝剞剟剡剣剤剥剧剩剪剫剬剭剮副剰割剳剴創剷剸剹剺剻剼剽剿劀劁劂劃劄劇劈劉劊劋劌劍劐劑劓劔劖劗劘劙力劝办功加务劢劣劦动助努劫劬劭励劲劳労劵効劼劾势勁勃勅勇勉勋勌勍勐勑勒勔動勖勗勘務勚勛勝勞募勢勣勤勦勧勩勰勱勲勳勴勵勷勸勺勻勾勿匀匁匂匃匄包匆匈匊匋匍匏匐匕化北匙匜匝匟匠匡匢匣匦匪匭匮匯匰匱匲匴匷匸匹区医匼匽匾匿區十千卅升午卉半卋卌卍华协卐卑卒卓協单卖南単博卜卞卟占卡卢卣卤卦卧卫卬卮卯印危卲即却卵卷卸卹卺卻卼卽卿厀厂厄厅历厉压厌厍厎厒厓厔厕厖厗厘厙厚厜厝厞原厠厢厣厤厥厦厨厩厬厭厮厰厲厳厴厹去厾县叁参參叄叆叇又叉及友双反収发叔叕取受变叙叚叛叟叠叡叢口古句另叨叩只叫召叭叮可台叱史右叵叶号司叹叻叼叽吁吃各吆合吉吊吋同名后吏吐向吒吓吔吕吖吗吚君吝吞吟吠吡吣吥否吧吨吩吪含听吭吮启吰吱吲吳吴吵吶吷吸吹吻吼吽吾呀呂呃呆呇呈呉告呋呌呎呐呑呒呓呔呕呖呗员呙呛呜呢呣呤呥呦周呪呫呬呰呱呲味呴呵呶呷呸呺呻呼命呾呿咀咁咂咄咅咆咇咈咉咊咋和咍咎咏咐咑咒咔咕咖咗咘咙咚咛咝咟咠咡咢咣咤咥咦咧咨咩咪咫咬咭咮咯咱咲咳咴咶咷咸咺咻咼咽咾咿哀品哂哃哄哅哆哇哈哉哌响哎哏哐哑哒哓哔哕哖哗哙哚哝哞哟員哢哤哥哦哧哨哩哪哫哭哮哱哲哳哶哺哼哽哿唁唃唄唅唆唇唈唉唌唏唐唑唒唔唕唖唗唘唚唛唝唠唢唣唤唦唧唪唫唬唭售唯唰唱唲唳唴唵唶唷唸唻唼唾唿啁啃啄啅商啉啊啍啎問啐啑啒啓啕啖啗啚啜啞啟啡啢啣啤啥啦啧啪啫啬啭啮啯啰啱啲啴啵啶啷啸啻啼啽啾喀喁喂喃善喆喇喈喉喊喋喌喍喏喑喒喓喔喘喙喚喜喝喟喢喣喤喥喦喧喨喩喪喫喬喭單喰喱喲喳喵営喷喹喻喼喽喾喿嗀嗁嗃嗄嗅嗆嗇嗈嗉嗊嗋嗌嗍嗎嗏嗐嗑嗒嗓嗔嗕嗖嗚嗛嗜嗝嗞嗟嗡嗢嗣嗤嗥嗦嗧嗨嗩嗪嗫嗬嗮嗯嗲嗳嗵嗶嗷嗹嗺嗼嗽嗾嗿嘀嘁嘂嘄嘅嘆嘈嘉嘌嘍嘎嘏嘐嘑嘒嘓嘔嘕嘖嘗嘘嘚嘛嘜嘝嘞嘟嘠嘡嘢嘣嘤嘥嘧嘩嘬嘭嘮嘯嘰嘱嘲嘳嘴嘵嘶嘷嘸嘹嘻嘽嘾嘿噀噁噂噅噆噇噈噉噊噌噍噎噏噓噔噗噘噙噚噛噜噝噞噠噢噣噤噥噦器噩噪噫噬噭噮噯噰噱噲噳噴噶噸噹噺噻噼嚀嚃嚄嚅嚆嚇嚈嚌嚍嚎嚏嚐嚒嚓嚕嚗嚘嚙嚚嚜嚞嚟嚢嚣嚤嚥嚦嚧嚨嚩嚪嚫嚬嚭嚮嚯嚲嚳嚴嚵嚶嚷嚼嚽嚾嚿囀囁囂囃囅囆囈囉囊囋囌囍囑囒囓囔囗囙囚四囝回囟因囡团団囤囧囪囫园囮囯困囱囲図围囵囶囷囹固国图囿圀圁圂圃圄圅圆圇圈圉圊國圌圍圏圐園圓圖圗團圙圜圞土圠圢圣圧在圩圪圫圬圭圮圯地圲圳圹场圻圾址坁坂坅均坉坊坋坌坍坎坏坐坑坒坖块坚坛坜坝坞坟坠坡坣坤坥坦坨坩坪坫坬坭坮坯坰坱坲坳坴坵坶坷坻坼坽坿垀垂垃垄垆垈型垌垍垎垏垒垓垔垕垗垘垙垚垛垜垝垞垟垠垡垢垣垤垦垧垩垫垭垮垯垰垱垲垴垵垸垺垻垼垽垾垿埂埃埅埆埇埈埋埌城埏埒埔埕埗埘埙埚埜埝域埠埡埢埤埧埨埪埫埭埮埯埳埴埵埶執埸培基埻埼埽堀堂堃堅堆堇堈堉堊堋堌堍堎堐堑堔堕堖堘堙堝堞堠堡堣堤堥堦堧堨堩堪堮堯堰報堲場堵堷堺堼堽堾堿塀塁塄塅塆塈塉塊塋塌塍塏塑塒塓塔塕塗塘塙塚塜塝塞塟塡塢塤塥塨塩填塬塭塯塰塱塲塴塵塸塹塺塼塽塾塿墀墁境墅墆墈墉墊墋墎墐墒墓墕墖増墘墙墚墜墝增墟墠墡墣墦墨墩墪墫墬墮墯墱墳墶墺墻墼墽墾壁壄壅壆壇壈壊壋壌壎壐壑壓壔壕壖壘壙壚壛壝壞壟壠壢壤壩壪士壬壮壯声壱売壳壴壶壷壸壹壺壻壼壽壾壿夀夃处夆备変夊夋夌复夎夏夐夒夔夕外夗夘夙多夜够夠夢夤夥大天太夫夬夭央夯失头夷夸夹夺夼夾奀奁奂奄奅奇奈奉奊奋奎奏奐契奓奔奕奖套奘奚奠奡奢奣奥奧奨奩奪奫奬奭奮奰奲女奴奶奸她奼好妀妁如妃妄妆妇妈妊妍妎妏妐妑妒妓妖妗妘妙妝妞妠妢妣妤妥妦妧妨妩妪妫妬妭妮妯妲妳妵妷妸妹妺妻妾姀姁姃姆姈姉姊始姌姍姎姏姐姑姒姓委姗姘姙姚姜姝姞姡姣姤姥姦姨姪姫姬姮姱姵姶姸姹姺姻姼姽姿娀威娃娄娅娆娇娈娉娋娌娍娎娑娓娖娘娙娛娜娟娠娣娥娩娫娭娯娱娲娳娴娵娶娷娸娼娽娿婀婁婆婉婊婌婍婐婑婓婕婗婘婚婞婟婠婢婣婤婥婦婧婪婫婬婭婯婳婴婵婶婷婸婺婻婼婿媂媃媄媆媊媐媒媓媔媕媖媗媚媛媜媞媟媢媥媦媧媪媬媭媮媯媰媱媲媳媵媸媺媻媼媽媾媿嫁嫂嫄嫇嫈嫉嫋嫌嫐嫒嫔嫕嫖嫗嫘嫚嫛嫜嫠嫡嫢嫣嫥嫦嫩嫪嫫嫭嫮嫰嫱嫲嫳嫴嫵嫶嫷嫺嫻嫽嫿嬀嬃嬅嬈嬉嬋嬌嬎嬐嬓嬖嬗嬙嬛嬝嬡嬢嬤嬥嬨嬪嬬嬭嬰嬲嬴嬷嬸嬽嬾嬿孀孁孃孅孆孈孋孌子孑孓孔孕孖字存孙孚孛孜孝孟孢季孤孥学孩孪孫孬孮孰孱孲孳孵孷學孺孻孼孽孿宁宂它宄宅宇守安宋完宍宎宏宓宕宗官宙定宛宜宝实実宠审客宣室宥宦宧宨宪宫宬宮宰害宴宵家宸容宽宾宿寀寁寂寃寄寅密寇寈寊寋富寍寐寑寒寓寔寕寖寗寘寙寛寜寝寞察寠寡寢寤寥實寧寨審寪寫寬寮寯寰寱寳寵寶寸对寺寻导対寿封専尃射尅将將專尉尊尋尌對導小尐少尒尓尔尕尖尗尘尙尚尜尝尟尢尤尥尧尨尪尬尭尰就尳尴尷尸尹尺尻尼尽尾尿局屁层屃屄居屆屇屈屉届屋屌屍屎屏屐屑屓屔展屖屘屙屚屜屝属屠屡屢屣層履屦屧屨屩屬屭屮屯山屴屹屺屻屼屾屿岀岁岂岅岈岉岊岋岌岍岏岐岑岒岔岖岗岘岙岚岛岜岝岞岟岠岡岢岣岤岧岨岩岪岫岬岭岮岯岰岱岳岵岷岸岻岽岿峁峂峃峄峅峇峉峊峋峌峍峐峒峗峘峙峚峛峝峞峠峡峣峤峥峦峧峨峩峪峬峭峮峯峰峱峴島峷峻峼峽峿崀崁崂崃崄崆崇崈崋崌崍崎崏崐崑崒崔崖崗崘崙崚崛崝崞崟崠崡崢崣崤崥崦崧崨崩崬崭崮崰崱崳崴崵崶崷崺崽崾崿嵁嵂嵃嵅嵇嵊嵋嵌嵎嵐嵑嵒嵕嵖嵗嵘嵙嵚嵛嵜嵝嵞嵣嵥嵨嵩嵫嵬嵮嵯嵰嵱嵲嵴嵷嵺嵻嵽嵾嵿嶀嶁嶂嶃嶄嶅嶆嶇嶈嶉嶊嶋嶌嶍嶒嶓嶔嶕嶗嶙嶚嶜嶝嶞嶟嶠嶡嶢嶦嶧嶨嶩嶪嶬嶭嶮嶰嶱嶲嶴嶵嶷嶸嶺嶼嶽巀巂巃巄巅巆巇巉巋巌巍巎巏巑巒巔巕巖巗巘巛川州巟巡巢巣工左巧巨巩巫差巯巰己已巳巴巵巶巷巹巻巽巾巿帀币市布帅帆师帊希帏帐帑帔帕帖帗帘帙帚帛帜帝帟帠帡帢帣帤帥带帧帨帩師帬席帮帯帰帱帳帴帶帷常帻帼帽帾幀幂幃幄幅幇幌幎幏幑幓幔幕幖幗幘幙幚幛幝幞幟幠幡幢幣幦幧幨幩幪幫幬幭幰干平年幵并幷幸幹幺幻幼幽幾广庀庁庂広庄庆庇庈庉床庋庌序庐庑库应底庖店庙庚庛府庞废庠庢庣庤庥度座庨庪庫庬庭庮庰庱庲庳庴庵庶康庸庹庻庼庽庾廁廂廃廄廅廆廇廈廉廊廋廌廍廎廐廑廒廓廔廕廖廗廘廙廚廛廜廝廞廟廠廡廢廣廥廦廧廨廩廪廬廮廯廰廱廳廵延廷廸廹建廻廼廾廿开弁异弃弄弅弆弇弈弉弊弋弌弍式弐弑弒弓弔引弖弗弘弚弛弝弟张弢弣弤弥弦弧弨弩弭弮弯弱弳張弶強弸弹强弼弾彀彃彄彅彆彈彉彊彋彌彎彏彐归当彔录彖彗彘彙彜彝彞彟彠彡形彣彤彥彦彧彩彪彫彬彭彯彰影彳彴彷彸役彻彼彾彿往征徂徃径待徇很徉徊律後徐徑徒従徕得徘徙徛徜從徠御徥徦徧徨復循徫徬徭微徯徲徳徴徵德徸徹徼徽徾徿忀忁心必忆忉忌忍忏忐忑忒忔忕忖志忘忙応忝忞忠忡忣忤忥忧忨忪快忬忭忮忯忱忳念忷忸忺忻忼忽忾忿怀态怂怃怄怅怆怊怋怍怎怏怐怒怓怔怕怖怗怙怚怛怜思怞怠怡怢急怦性怨怩怪怫怬怭怮怯怱怲怳怴怵怷怹总怼怿恀恁恂恃恅恆恇恉恊恋恌恍恐恒恓恔恕恙恚恛恝恞恟恠恡恢恣恤恥恧恨恩恪恫恬恭息恰恲恳恵恶恸恹恺恻恼恽恿悀悁悃悄悅悆悇悈悉悊悌悍悐悒悔悕悖悗悚悛悝悞悟悠悢患悤悦悧您悩悪悫悬悭悯悰悱悲悳悴悵悶悷悸悹悺悻悼悽悾惄情惆惇惈惉惊惋惌惎惏惑惓惔惕惘惙惚惛惜惝惟惠惡惢惣惤惥惦惧惨惩惪惫惬惭惮惯惰惱惲想惴惵惶惷惸惹惺惻惼惾愀愁愃愅愆愈愉愊愍愎意愐愒愓愔愕愖愚愛愜感愠愣愤愦愧愨愫愬愭愮愯愲愴愶愷愼愽愾愿慁慂慄慅慆慇慈慉慊態慌慍慎慑慒慓慔慕慘慚慛慜慝慞慟慢慣慤慥慧慨慪慫慬慭慮慰慱慲慳慴慵慶慷慹慺慼慽慾憀憂憃憇憉憊憋憍憎憐憑憒憓憔憕憖憙憚憛憝憟憡憢憤憧憨憩憪憫憬憭憮憯憰憱憲憴憵憶憷憸憺憻憼憾憿懁懂懃懅懆懇懈應懊懋懌懍懐懑懒懔懘懞懟懠懣懤懥懦懧懨懪懫懭懮懰懱懲懵懶懷懸懺懻懼懽懾懿戀戁戃戄戆戇戈戉戊戋戌戍戎戏成我戒戔戕或戗战戙戚戛戞戟戠戡戢戣戤戥戦戧戩截戬戭戮戯戰戱戲戳戴戶户戸戹戺戻戼戽戾房所扁扂扃扅扆扇扈扉扊手才扎扐扑扒打扔払托扙扛扜扞扡扢扣扤扥扦执扩扪扫扬扭扮扯扰扱扳扶批扺扻扼扽找承技抃抄抆抇抉把抌抎抏抑抒抓抔投抖抗折抚抛抜抝択抟抠抡抢护报抦抨抪披抬抭抮抱抴抵抶抹抻押抽抾抿拂拃拄担拆拇拈拉拊拋拌拍拎拏拐拑拒拓拔拕拖拗拘拙拚招拜拝拟拠拡拢拣拤拥拦拧拨择拫括拭拮拯拱拲拳拴拵拶拷拸拹拺拼拽拾拿挀持挂挃指挈按挋挌挍挎挏挐挑挓挖挙挚挛挝挞挟挠挡挣挤挥挦挨挩挪挫挬振挱挲挴挵挶挹挺挻挼挽挾挿捁捂捃捄捅捆捉捊捋捌捍捎捏捐捒捔捕捖捗捘捜捞损捡换捣捥捦捧捨捩捫捭据捯捱捲捵捶捷捺捻捼捽掀掂掃掄掇授掉掊掌掎掏掐排掔掖掗掘掙掛掜掝掞掟掠採探掣掤接控推掩措掫掬掭掮掯掰掱掲掳掴掷掸掺掻掼掽掾揀揃揄揅揆揉揊揍揎描提揑插揔揕揖揘揚換揜揝揟揠握揣揤揥揨揩揪揫揭揮揯揱揲揳援揵揶揷揸揹揺揽揿搀搁搂搅搆搉搊搋搌損搏搐搒搓搔搕搖搗搘搚搛搜搞搟搠搡搢搣搤搥搦搧搨搪搫搬搭搯搰搳搴搵搶搷搹携搽搾摀摁摂摃摄摅摆摇摈摊摋摍摎摏摐摑摒摓摔摘摛摜摝摞摟摠摡摦摧摩摭摮摯摰摲摳摴摵摶摸摹摺摻摽摿撂撃撄撅撇撈撊撋撌撏撐撑撒撓撕撖撗撙撚撜撝撞撟撠撢撣撤撥撦撩撫撬播撮撰撲撳撵撷撸撹撺撻撼撽撾撿擀擁擂擄擅擇擉擊擋操擎擏擐擒擓擔擕擖擗擘據擞擠擡擢擣擤擥擦擧擨擩擫擬擭擯擰擱擲擴擷擸擺擻擼擽擾擿攀攄攆攇攉攌攍攏攐攒攓攔攕攖攗攘攙攛攜攝攞攠攡攢攣攤攥攦攩攪攫攬攭攮支攲攳攴收攷攸改攻攽放政敁敂故敆效敉敊敌敍敎敏救敓敔敕敖敗敘教敛敜敝敞敟敢散敤敦敧敩敪敫敬敭敯数敲敳整敵敶敷數敹敺敻敾敿斀斁斂斃斄文斉斋斌斎斐斑斒斓斔斕斖斗料斛斜斝斞斟斠斡斤斥斧斨斩斪斫斬断斮斯新斲斵斶斷方斻於施斿旁旂旃旄旅旆旉旋旌旍旎族旐旒旓旖旗旙旚旛旝旞旟无旡既日旦旧旨早旬旭旮旯旰旱旲旳旴旵时旷旸旺旻旼旽旾旿昀昂昃昄昆昇昈昉昊昌昍明昏昐昒易昔昕昙昚昛昜昝昞星映昡昢昣昤春昦昧昨昪昫昬昭是昰昱昲昳昴昵昶昺昼昽显晁時晃晅晇晈晉晊晋晌晏晐晒晓晔晕晖晗晙晚晛晜晝晞晟晡晢晤晥晦晧晨晩晪晫晬晭普景晰晱晳晴晶晷晸智晻晼晾暀暁暂暃暄暅暆暇暈暉暊暋暌暍暎暏暐暑暕暖暗暘暝暟暠暡暢暦暧暨暩暫暮暰暱暲暴暵暶暸暹暻暾暿曀曄曅曆曇曈曉曊曋曌曏曒曔曖曙曚曛曜曝曠曡曣曦曧曨曩曬曭曮曰曲曳更曶曷書曹曺曼曽曾替最朁會朄朅月有朊朋服朏朐朒朓朔朕朖朗朘望朝朞期朡朢朣朦朧木未末本札朮术朱朳朴朵朶朸朹机朻朼朽朾朿杀杁杂权杄杅杆杇杈杉杋杌李杏材村杓杕杖杗杙杜杝杞束杠条杢杣杤来杧杨杩杪杬杭杮杯杰東杲杳杴杵杶杷杻杼松板极构枅枆枇枉枋枌枍枎枏析枒枓枕枖林枘枚枛果枝枞枠枡枢枣枥枧枨枪枫枭枯枰枱枲枳枴枵架枷枸枹枺枻枼柀柁柂柃柄柅柈柉柊柍柎柏某柑柒染柔柖柘柙柚柜柝柞柟柠柢柣柤查柧柩柫柬柮柯柰柱柲柳柴柵柶柷柸柹柺査柽柾柿栀栁栂栃栄栅标栈栉栊栋栌栎栏栐树栒栓栔栖栗栘栚栜栝栞栟校栢栨栩株栫栭栯栰栱栲栳栴栵样核根栻格栽栾桀桁桂桃桄桅框案桉桊桋桌桎桏桐桑桒桓桔桕桖桙桜桝桟桠桡桢档桤桥桦桧桨桩桫桭桮桯桱桲桴桵桶桷桹桺桼桾桿梀梁梂梃梅梆梇梉梌梏梐梒梓梔梗梘梛梜條梟梠梡梢梣梦梧梨梩梫梬梭梮梯械梱梲梳梴梵梶梼梽梾梿检棁棂棃棄棆棉棊棋棌棍棐棒棓棕棖棗棘棚棛棜棟棠棡棣棤棧棨棩棪棫棬森棯棰棱棲棳棵棶棷棸棹棺棻棼棽椀椁椄椅椆椇椈椉椋椌植椎椏椐椑椒椓椔椗椙検椟椠椤椥椪椭椰椲椳椴椵椶椸椹椽椿楀楂楄楅楈楉楊楎楒楓楔楕楖楗楘楙楚楛楝楞楟楠楡楢楣楤楥楦楧楨楩楪楫楬業楮楯楰楱楳極楶楷楸楹楺楼楽楿概榃榄榅榆榇榈榉榊榍榎榑榔榕榖榗榘榙榛榜榞榠榢榣榤榦榧榨榩榪榫榬榭榮榯榰榱榲榴榶榷榹榻榼榽榾榿槀槁槃槄槇槉槊構槌槍槎槐槑槓槔槕槗様槙槚槛槜槟槠槢槤槥槧槨槩槬槭槮槰槱槲槳槷槸槺槻槼槽槾槿樀樁樂樅樆樊樋樑樓樔樕樗樘標樛樝樞樟樠模樣樧樨権横樫樯樱樲樴樵樸樹樺樻樽樾樿橀橁橄橇橈橉橋橐橑橒橓橘橙橚橛橜橝橞機橡橢橤橥橦橧橨橪橫橭橱橶橹橼橾橿檀檁檃檄檇檉檊檍檎檐檑檓檔檕檖檗檛檜檝檞檟檠檡檢檣檥檦檨檩檫檬檮檯檳檵檸檹檻檽櫂櫃櫅櫆櫈櫋櫌櫐櫑櫓櫙櫚櫛櫜櫝櫞櫟櫠櫡櫥櫧櫨櫪櫫櫬櫰櫱櫳櫸櫹櫺櫻櫼櫾櫿欂欃欄欅欈欉權欋欏欐欑欒欖欘欞欠次欢欣欤欥欧欨欬欭欯欱欲欳欴欵欶欷欸欹欺欻欽款欿歂歃歅歆歇歈歉歊歋歌歍歎歐歓歔歕歙歛歜歟歠歡止正此步武歧歩歪歭歯歲歳歴歶歷歸歹死歼歾歿殀殁殂殃殄殆殇殈殉殊残殍殏殑殒殓殔殖殗殘殙殚殛殞殟殠殡殢殣殤殥殦殧殪殫殭殮殯殰殲殳殴段殶殷殺殻殼殽殿毀毁毂毃毄毅毆毇毈毉毊毋毌母毎每毐毒毓比毕毖毗毘毙毚毛毞毡毣毤毦毨毪毫毬毯毰毲毳毴毵毸毹毻毼毽毾毿氀氁氂氄氅氆氇氈氊氌氍氏氐民氓气氕氖気氘氙氚氛氟氠氡氢氣氤氦氧氨氩氪氫氬氮氯氰氲氳水氶氷永氹氻氽氾氿汀汁求汃汆汇汈汉汊汋汍汎汏汐汒汔汕汗汙汚汛汜汝汞江池污汤汦汧汨汩汪汫汭汯汰汱汲汳汴汶汸汹決汽汾沁沂沃沄沅沆沇沈沉沋沌沍沏沐沒沓沔沕沖沘沙沚沛沜沟没沢沣沤沥沦沧沨沩沪沫沬沭沮沯沱河沴沶沷沸油沺治沼沽沾沿況泂泃泄泅泆泇泉泊泌泐泒泓泔法泖泗泙泚泛泜泝泞泠泡波泣泥泧注泩泪泫泬泭泮泯泰泱泲泳泵泷泸泺泻泼泽泾洀洁洄洇洈洊洋洌洍洎洑洒洓洗洘洙洚洛洞洟洢洣洤津洧洨洩洪洫洭洮洱洲洳洴洵洶洸洹洺活洼洽派洿流浃浄浅浆浇浈浉浊测浍济浏浐浑浒浓浔浕浘浙浚浛浜浞浟浠浡浢浣浤浥浦浧浩浪浬浭浮浯浰浲浴浵浶海浸浹浺浼浿涂涄涅涆涇消涉涊涌涍涎涐涑涒涓涔涕涖涗涘涙涛涝涞涟涠涡涢涣涤润涧涨涩涪涫涬涮涯液涳涴涵涷涸涼涽涾涿淀淄淅淆淇淈淊淋淌淍淎淏淑淒淓淔淖淘淙淚淛淜淝淞淟淠淡淢淤淥淦淨淩淪淫淬淭淮淯淰深淳淴淵淶混淸淹淺添淼淽渀渃清渇済渉渊渋渌渍渎渐渑渓渔渕渗渙渚減渜渝渟渠渡渢渣渤渥渦渨温渫測渭渮港渰渱渲渳渴游渹渺渻渼渽渾渿湀湁湃湄湅湆湇湉湊湋湍湎湑湓湔湖湘湛湜湝湞湟湠湡湢湣湥湧湨湩湫湮湯湱湲湳湴湻湼湾湿満溁溃溅溆溇溈溉溋溍溎溏源溓溔準溘溙溚溛溜溝溞溟溠溢溥溦溧溪溫溮溯溰溱溲溳溴溵溶溷溹溺溻溼溽溾滀滁滂滃滄滅滆滇滈滉滋滌滍滎滏滑滒滓滔滕滖滗滘滙滚滛滜滝滞滟滠满滢滤滥滦滧滨滩滪滫滬滭滮滯滱滲滴滵滷滸滹滻滽滾滿漁漂漃漅漆漇漈漉漊漋漎漏漑漒漓演漕漖漘漙漚漠漢漣漤漥漦漧漩漪漫漬漭漮漯漰漱漲漳漴漵漶漷漸漹漻漼漾漿潀潁潃潄潅潆潇潋潍潎潏潐潑潒潔潕潖潗潘潙潚潛潜潝潞潟潠潢潤潦潩潬潭潮潯潰潲潳潴潵潶潷潸潺潼潽潾潿澀澁澂澄澅澆澇澈澉澋澌澍澎澐澒澓澔澖澗澛澜澞澠澡澣澤澥澦澧澨澩澪澫澭澮澯澱澳澴澶澹澻澼澽激濁濂濃濄濆濇濈濉濊濋濌濎濑濒濔濕濘濙濛濜濞濟濠濡濣濤濦濧濩濫濬濭濮濯濰濱濲濴濶濺濻濼濾濿瀀瀁瀄瀅瀆瀇瀉瀋瀌瀍瀎瀏瀑瀔瀕瀖瀘瀙瀚瀛瀝瀞瀟瀠瀡瀢瀣瀦瀧瀨瀩瀫瀬瀯瀰瀱瀲瀴瀵瀶瀷瀸瀹瀺瀼瀾瀿灂灃灄灈灉灊灋灌灏灑灕灖灗灘灛灝灞灟灡灣灤灧灨灩灪火灭灯灰灴灵灶灸灼災灾灿炀炁炅炆炉炊炌炎炒炔炕炖炘炙炜炝炟炣炤炫炬炭炮炯炰炱炲炳炷炸点為炻炼炽炾烀烁烂烃烈烉烊烋烏烑烒烓烔烖烘烙烛烜烝烟烠烤烦烧烨烩烫烬热烯烰烱烴烶烷烹烺烻烽焃焄焆焉焊焌焍焐焓焔焕焖焗焘焙焚焜焞焠無焣焦焮焯焰焱焴然焻焼焿煁煃煅煆煇煉煊煋煌煎煐煑煒煓煔煕煖煗煙煚煜煝煞煟煠煢煣煤煥煦照煨煩煬煮煲煳煴煵煶煸煺煽熀熂熄熅熇熈熉熊熏熐熒熔熖熗熘熙熚熛熜熟熠熤熥熨熬熯熰熱熲熳熵熸熹熺熻熼熾熿燀燁燂燃燄燅燈燉燊燋燎燏燐燒燔燕燖燗燘燙燚燜燝營燠燡燥燦燧燫燬燭燮燴燹燻燼燾燿爀爁爂爆爇爊爌爍爐爓爔爕爚爛爝爞爟爢爣爨爪爬爭爯爰爱爲爵父爷爸爹爺爻爼爽爾爿牀牁牂牄牆片版牉牋牌牍牎牏牒牓牕牖牘牙牚牛牝牟牠牡牢牣牤牥牦牧物牬牮牯牰牲牴牵牷牸特牺牻牼牽牾牿犀犁犂犄犅犆犇犈犉犊犋犌犍犎犏犐犑犒犓犕犖犗犘犚犛犝犞犟犠犢犣犤犥犦犧犨犩犪犬犮犯犰犴犵状犷犸犹犽犿狀狁狂狃狄狆狈狉狊狋狌狍狎狐狑狒狓狔狖狗狘狙狚狛狜狝狞狟狠狡狣狤狥狦狨狩狪狫独狭狮狯狰狱狲狳狴狶狷狸狹狺狻狼狽狾狿猀猁猂猃猄猇猈猊猋猎猏猑猒猓猕猖猗猘猙猛猜猝猞猟猡猢猣猥猦猧猨猩猪猫猬猭献猯猰猱猲猳猴猵猶猷猹猺猻猼猾猿獀獁獂獃獄獅獇獊獋獌獍獎獏獐獑獒獗獘獛獝獞獟獠獡獢獣獥獦獧獨獩獪獫獬獭獮獯獰獲獳獴獵獶獷獸獺獻獼獽獾獿玀玁玂玃玄玅玆率玈玉玊王玎玏玑玒玓玔玕玖玗玘玙玚玛玞玟玠玡玢玤玥玦玧玨玩玫玭玮环现玱玲玳玶玷玹玺玻玼玾玿珀珂珅珇珈珉珊珋珌珍珎珏珐珑珒珓珕珖珙珛珝珞珠珡珢珣珤珥珦珧珩珪珫班珮珰珲珵珶珷珸珹珺珽現琀琁球琄琅理琇琈琉琊琍琎琏琐琔琖琚琛琟琡琢琤琥琦琨琪琫琬琭琮琯琰琱琲琳琴琵琶琹琺琼琿瑀瑁瑂瑃瑄瑅瑆瑇瑈瑊瑋瑑瑒瑓瑔瑕瑖瑗瑙瑚瑛瑜瑝瑞瑟瑠瑢瑣瑤瑥瑧瑨瑩瑪瑬瑭瑯瑰瑱瑲瑳瑴瑵瑶瑷瑺瑽瑾瑿璀璁璃璄璅璆璇璈璉璊璋璌璎璐璒璕璗璘璙璚璜璞璟璠璡璢璣璥璦璧璨璩璪璫璬璮環璱璲璵璸璹璺璽璿瓀瓁瓅瓈瓊瓋瓌瓏瓑瓒瓔瓖瓘瓚瓛瓜瓝瓞瓟瓠瓡瓢瓣瓤瓥瓦瓨瓩瓬瓮瓯瓴瓵瓶瓷瓻瓽瓾瓿甀甂甃甄甈甋甌甍甏甐甑甒甓甔甕甖甗甘甙甚甜甝甞生甡產産甥甦用甩甪甫甬甭甮甯田由甲申电男甸甹町画甽甾甿畀畅畇畈畊畋界畎畏畐畑畔畖留畚畛畜畝畟畠畢畣畤略畦畧番畫畬畯異畱畲畳畴畵當畷畸畹畺畽畿疀疁疃疆疇疊疋疌疍疎疏疐疑疒疔疕疖疗疘疙疚疝疟疠疡疢疣疤疥疧疫疬疭疮疯疰疱疲疳疴疵疸疹疻疼疽疾疿痀痁痂痃痄病症痈痉痊痋痌痍痎痏痐痑痒痓痔痕痗痘痙痚痛痝痞痟痠痡痢痣痤痦痧痨痩痪痫痯痰痱痲痳痴痷痸痹痺痻痼痾痿瘀瘁瘃瘅瘆瘈瘉瘊瘋瘌瘍瘏瘐瘑瘓瘕瘖瘗瘘瘙瘛瘜瘝瘞瘟瘠瘡瘢瘣瘤瘥瘦瘧瘩瘪瘫瘭瘮瘯瘰瘱瘲瘳瘴瘵瘸瘺瘻瘼瘽瘾瘿癀療癃癄癅癆癇癈癉癌癍癐癒癓癔癖癗癘癙癜癞癟癠癡癢癣癤癥癧癩癪癫癬癭癮癯癰癱癲癵癸癹発登發白百癿皁皂的皆皇皈皊皋皎皏皐皑皓皕皖皙皚皛皜皝皞皤皦皫皭皮皯皰皱皲皴皵皷皸皺皻皽皾皿盂盃盄盅盆盇盈盉益盋盌盍盎盏盐监盒盓盔盖盗盘盛盜盝盞盟盡監盤盥盦盧盨盩盪盫盬盭目盯盱盲盳直盷相盹盺盻盼盾眀省眃眄眅眇眈眉眊看県眍眐眑眒眓眕眙眚眛眜眝眞真眠眢眣眥眦眨眩眬眭眯眱眲眳眴眵眶眷眸眹眺眼眽眾着睁睃睄睅睆睇睊睋睌睍睎睏睐睑睒睔睕睖睚睛睜睞睟睠睡睢督睥睦睧睨睩睪睫睬睭睮睯睹睺睼睽睾睿瞀瞁瞂瞄瞅瞇瞉瞋瞌瞍瞎瞏瞑瞒瞓瞖瞘瞚瞛瞜瞝瞞瞟瞠瞡瞢瞥瞧瞩瞪瞫瞬瞭瞰瞲瞳瞴瞵瞶瞷瞻瞼瞽瞿矁矂矇矉矊矌矍矎矏矐矓矔矕矗矘矙矚矛矜矞矠矢矣知矧矩矫矬短矮矯矰矱矲石矴矶矸矹矺矻矼矽矾矿砀码砂砄砅砆砉砋砌砍砎砏砐砑砒研砕砖砗砘砚砜砝砟砠砢砣砥砦砧砨砩砫砬砭砮砯砰砲砳破砵砷砸砹砺砻砼砾础硁硂硃硅硇硈硉硊硌硍硎硏硐硒硓硔硕硖硗硙硚硜硝硠硤硥硨硩硪硫硬硭确硯硰硱硼硾硿碁碃碄碅碆碇碈碉碌碍碎碏碑碓碔碕碗碘碚碛碜碞碟碡碣碥碧碨碩碪碫碭碰碱碲碳碴碶碸碹確碻碼碽碾磁磃磅磈磉磊磋磌磍磎磏磐磑磔磕磘磙磚磛磜磝磞磟磡磣磥磧磨磩磪磬磯磱磲磳磴磵磷磹磺磻磼磽磾磿礁礂礄礅礉礌礎礐礑礒礓礔礙礛礜礝礞礠礣礥礦礧礨礪礫礬礭礮礱礴礵示礼礽社礿祀祁祂祃祅祆祇祈祉祊祋祌祎祏祐祑祒祓祔祕祖祗祙祚祛祜祝神祟祠祡祢祣祤祥祧票祩祪祫祭祯祲祳祴祷祸祹祺祼祾祿禀禁禂禃禄禅禇禊禋禍禎福禑禒禓禔禕禖禗禘禚禛禜禟禠禡禤禦禧禨禩禪禫禬禭禮禰禱禳禴禵禷禸禹禺离禼禽禾禿秀私秃秅秆秈秉秊秋种秎秏科秒秔秕秖秘租秠秣秤秦秧秩秪秫秬秭秮积称秳秶秸秺移秽秾稀稂稃稅稆稈稉稊程稌稍税稑稒稔稗稘稙稚稜稞稟稠稣稫稬稭種稯稰稱稲稳稷稹稺稻稼稽稾稿穀穂穄穅穆穇穈穉穊穋穌積穎穏穑穖穗穙穛穜穟穠穡穢穣穤穧穨穩穫穬穭穮穰穱穴穵究穷穸穹空穽穾穿窀突窃窄窅窆窈窉窊窋窌窍窎窏窐窑窒窓窔窕窖窗窘窙窜窝窞窟窠窢窣窥窦窨窩窪窫窬窭窮窯窰窱窲窳窴窵窶窸窺窻窾窿竀竁竃竄竅竇竈竊立竑竖竘站竚竜竝竞竟章竢竣童竦竩竪竫竭端竴競竷竹竺竻竽竿笃笄笅笆笈笊笋笏笐笑笓笔笕笘笙笛笝笞笠笢笤笥符笨笩笪笫第笭笮笯笰笱笲笳笴笵笸笹笺笻笼笾筀筄筅筆筇筈等筊筋筌筍筏筐筑筒答策筘筚筛筜筝筞筠筡筢筤筥筦筧筩筬筭筮筯筰筱筲筳筴筵筶筷筹筻筼签简箄箅箆箇箈箊箋箌箍箎箏箐箑箒箓箔箕箖算箘箙箛箜箝箠管箢箦箧箨箩箪箫箬箭箯箱箴箷箸箹箾節篁範篆篇築篊篋篌篎篑篓篔篕篙篚篛篝篞篠篡篣篤篥篦篧篨篩篪篭篮篯篱篲篳篴篷篸篹篻篼篽篾篿簀簁簂簃簅簇簉簋簌簍簎簏簑簒簕簖簗簙簜簝簞簟簠簡簢簣簥簦簧簨簩簪簫簬簭簮簰簳簷簸簹簺簻簼簽簾簿籀籁籃籈籉籊籌籍籐籔籗籙籛籜籝籟籠籣籤籥籦籧籩籪籬籮籯籲米籴籸籹籺类籼籽籾粂粃粄粇粉粊粋粍粑粒粔粕粗粘粛粜粝粞粟粢粤粥粦粧粩粪粮粱粲粳粵粹粺粻粼粽精粿糀糁糅糇糈糉糊糌糍糎糒糔糕糖糗糙糜糝糞糟糠糢糧糨糪糬糯糰糱糲糴糵糶糷糸糹糺系糾紀紂紃約紅紆紇紈紉紊紋納紎紐紑紒紓純紕紖紗紘紙級紛紜紝紞紟素紡索紥紧紩紫紬紮累細紱紲紳紵紶紸紹紺紼紽紾紿絀絁終絃組絅絆絇経絎絏結絓絕絖絘絛絜絝絞絡絢絣給絧絨絪絫絭絮絯絰統絲絳絵絶絷絹絺絻絼絽絿綀綁綃綄綅綆綈綉綌綍綎綏綑經綔綖継続綜綝綞綟綠綢綣綦綧綪綫綬維綮綯綰綱網綳綴綵綷綸綹綺綻綼綽綾綿緀緁緂緃緄緅緆緇緈緊緋緌緎総緐緑緒緔緖緗緘緙線緛緜緝緞締緡緣緤緥緦緧編緩緪緬緯緰緱緲緳練緵緶緷緹緺緻縁縂縃縄縅縈縉縊縋縌縎縏縐縑縒縓縕縗縚縛縜縝縞縟縠縡縢縣縤縦縧縩縪縫縭縮縯縰縱縲縳縴縵縶縷縸縹縻縼總績縿繀繁繂繃繄繅繆繇繈繉繊繋繍繐繑繒織繕繖繘繙繚繜繝繞繟繠繡繢繣繦繧繩繪繫繭繮繯繰繲繳繴繵繶繷繸繹繺繻繼繽繾纀纁纂纆纇纈纊纋續纍纏纐纑纒纓纔纕纖纗纘纚纛纜纠纡红纣纤纥约级纨纩纪纫纬纭纮纯纰纱纲纳纴纵纶纷纸纹纺纻纼纽纾线绀绁绂练组绅细织终绉绊绋绌绍绎经绐绑绒结绔绕绖绗绘给绚绛络绝绞统绠绡绢绣绤绥绦继绨绩绪绫续绮绯绰绱绲绳维绵绶绷绸绹绺绻综绽绾绿缀缁缂缃缄缅缆缇缈缉缊缌缎缐缑缒缓缔缕编缗缘缙缚缛缜缝缞缟缠缡缢缣缤缥缦缧缨缩缪缫缬缭缮缯缰缱缲缳缴缵缶缸缹缺缽缾缿罂罃罄罅罇罈罉罊罋罌罍罎罏罐网罔罕罗罘罚罛罜罝罞罟罠罡罢罣罥罦罧罨罩罪罫罬罭置罰罱署罳罴罵罶罷罹罺罻罼罽罾罿羁羂羃羅羆羇羈羉羊羋羌羍美羑羒羓羔羕羖羗羙羚羛羜羝羞羟羠羡羢羣群羥羦羧羨義羬羭羯羰羱羲羳羴羵羶羷羸羹羺羼羽羾羿翀翁翂翃翄翅翇翈翊翋翌翍翎翏翐翑習翔翕翘翙翚翛翜翟翠翡翢翣翥翦翨翩翪翫翬翭翮翯翰翱翲翳翴翷翹翺翻翼翽翾翿耀老考耄者耆耇耋而耍耎耏耐耑耒耔耕耖耗耘耙耜耞耟耠耡耢耤耥耦耧耨耩耪耬耮耰耱耳耴耵耶耷耸耹耻耽耾耿聂聃聆聇聈聊聋职聍聏聑聒联聖聘聚聝聞聟聡聨聩聪聯聰聱聲聳聴聵聶職聹聽聾聿肂肃肄肅肆肇肉肋肌肎肏肐肒肓肕肖肘肙肚肛肜肝肟肠股肢肣肤肥肧肩肪肫肭肮肯肱育肴肵肷肸肹肺肼肽肾肿胀胁胂胃胄胅胆胇胈胉胊背胍胎胏胐胑胔胕胖胗胘胙胚胛胜胝胞胠胡胣胤胥胧胨胩胪胫胬胭胯胰胱胲胳胴胵胶胷胸胹胺胻胼能胾脀脁脂脃脅脆脇脈脉脊脍脎脏脐脑脒脓脔脕脖脗脘脙脚脛脞脟脡脣脤脥脧脩脫脬脭脯脰脱脲脳脶脷脸脹脽脾脿腃腄腆腇腈腊腋腌腍腎腏腐腑腒腓腔腕腖腘腙腚腛腜腞腠腢腤腥腦腧腨腩腫腭腮腯腰腱腲腳腴腶腷腸腹腺腻腼腽腾腿膀膂膃膆膇膈膉膊膋膌膍膏膑膕膘膙膚膛膜膝膞膟膠膢膣膥膦膨膩膫膬膮膰膱膲膳膴膵膷膹膺膻膽膾膿臀臂臃臄臅臆臇臈臉臊臋臌臍臏臐臑臒臓臕臗臘臙臚臛臜臝臞臟臠臡臢臣臥臦臧臨臩自臬臭臮臯臲至致臷臸臺臻臼臾臿舀舁舂舄舅舆與興舉舊舋舌舍舎舐舑舒舔舕舖舗舘舛舜舝舞舟舠舡舢舣舥舨舩航舫般舭舯舰舱舲舳舴舵舶舷舸船舺舻舼舽舾艀艂艄艅艇艉艋艎艏艐艑艒艓艔艕艘艙艚艛艟艤艦艨艫艮良艰艱色艳艴艵艶艷艸艹艺艼艽艾艿芀节芃芄芅芈芊芋芍芎芏芐芑芒芓芔芗芘芙芚芛芜芝芞芟芠芡芣芤芥芦芧芨芩芪芫芬芭芮芯芰花芲芳芴芵芷芸芹芺芻芼芽芾芿苀苁苄苅苇苈苉苊苋苌苍苎苏苑苒苓苔苕苖苗苘苙苛苜苝苞苟苠苡苣苤若苦苧苨苪苫苬苯英苲苳苴苵苶苷苹苺苻苼苾茀茁茂范茄茅茆茇茈茉茋茌茍茎茏茑茓茔茕茖茗茘茙茚茛茜茝茠茢茤茥茦茧茨茩茪茫茬茭茮茯茱茲茳茴茵茶茷茸茹茺茻茼茽茿荀荁荂荃荄荅荆荇草荊荍荎荏荐荑荒荓荔荖荘荙荚荛荜荞荟荠荡荣荤荥荦荧荨荩荪荫荬荭荮药荳荴荵荷荸荺荻荼荽荾莁莃莅莆莉莊莋莌莍莎莏莐莒莓莔莕莖莘莙莛莜莝莞莠莢莣莤莥莧莨莩莪莫莬莰莱莲莳莴莵莶获莸莹莺莼莽莿菀菁菂菅菆菇菈菉菊菋菌菍菎菏菑菓菔菕菖菘菙菜菝菞菟菠菡菣菤菥菧菨菩菪菫華菰菱菲菳菴菵菶菸菹菺菼菽菾菿萁萃萄萆萇萉萊萋萌萍萎萏萐萑萒萘萚萜萝萠萡萣萤营萦萧萨萩萬萭萯萰萱萲萳萴萵萷萸萹萺萼落萿葀葂葃葅葆葉葊葋葌葍葎葑葒葖著葙葚葛葜葝葞葟葡董葤葥葦葧葩葫葬葭葯葰葱葳葴葵葶葷葸葹葺葽蒂蒄蒆蒇蒈蒉蒋蒌蒍蒎蒏蒐蒓蒔蒗蒘蒙蒚蒛蒜蒝蒞蒟蒠蒡蒢蒤蒧蒨蒩蒪蒫蒬蒭蒮蒯蒱蒲蒴蒶蒸蒹蒺蒻蒼蒽蒾蒿蓀蓁蓂蓄蓆蓇蓉蓊蓋蓌蓍蓎蓏蓐蓑蓓蓖蓘蓚蓝蓟蓠蓢蓣蓥蓦蓧蓨蓩蓪蓫蓬蓭蓮蓯蓰蓱蓲蓳蓴蓶蓷蓹蓺蓻蓼蓽蓾蓿蔀蔂蔃蔄蔆蔇蔈蔉蔊蔌蔍蔎蔏蔑蔒蔓蔔蔕蔖蔗蔘蔚蔜蔝蔞蔟蔠蔡蔣蔤蔥蔦蔧蔨蔩蔪蔫蔬蔭蔮蔰蔱蔴蔵蔷蔸蔹蔺蔻蔼蔽蔾蔿蕀蕁蕃蕄蕅蕆蕇蕈蕉蕊蕋蕍蕎蕑蕒蕓蕕蕖蕗蕘蕙蕚蕛蕝蕞蕟蕠蕡蕢蕣蕤蕦蕧蕨蕩蕪蕫蕭蕮蕰蕱蕲蕴蕵蕷蕸蕹蕺蕻蕼蕾蕿薀薁薂薃薄薅薆薇薈薉薊薋薌薍薎薏薐薑薔薕薖薗薘薙薚薛薜薝薞薟薠薡薢薣薤薦薧薨薩薪薫薬薭薮薯薰薱薲薳薴薵薶薷薸薹薺薽薾薿藀藁藂藃藄藅藆藇藈藉藋藍藎藏藐藑藒藓藔藕藗藘藙藚藜藝藟藠藡藣藤藥藦藨藩藪藫藬藭藰藱藲藴藶藷藸藹藺藻藼藾藿蘀蘁蘂蘄蘅蘆蘇蘉蘊蘋蘌蘐蘑蘓蘖蘗蘘蘙蘚蘛蘜蘞蘟蘠蘡蘢蘣蘤蘥蘦蘧蘩蘪蘬蘭蘮蘱蘳蘴蘵蘶蘸蘹蘺蘻蘼蘾蘿虀虃虆虇虈虉虋虌虎虏虐虑虒虓虔處虖虙虚虛虜虞號虡虢虣虤虧虨虩虪虫虬虭虮虯虰虱虴虵虷虸虹虺虻虼虽虾虿蚀蚁蚂蚄蚅蚆蚇蚊蚋蚌蚍蚑蚓蚕蚖蚗蚘蚙蚚蚜蚝蚞蚡蚢蚣蚤蚥蚧蚨蚩蚪蚬蚯蚰蚱蚲蚳蚴蚵蚶蚷蚸蚹蚺蚻蚼蚿蛀蛁蛂蛃蛄蛅蛆蛇蛈蛉蛊蛋蛌蛍蛎蛏蛐蛑蛓蛔蛕蛖蛗蛘蛙蛚蛛蛜蛝蛞蛟蛢蛣蛤蛦蛩蛪蛫蛬蛭蛮蛯蛰蛱蛲蛳蛴蛵蛶蛷蛸蛹蛺蛻蛾蜀蜁蜂蜃蜄蜆蜇蜈蜉蜊蜋蜌蜍蜎蜐蜑蜒蜓蜕蜖蜗蜘蜙蜚蜛蜜蜞蜠蜡蜢蜣蜤蜥蜦蜧蜨蜩蜪蜬蜭蜮蜯蜰蜱蜲蜳蜴蜵蜷蜸蜺蜻蜼蜾蜿蝀蝁蝂蝃蝆蝇蝈蝉蝋蝌蝍蝎蝐蝑蝒蝓蝔蝕蝖蝗蝘蝙蝚蝛蝜蝝蝞蝟蝠蝡蝣蝤蝥蝦蝨蝪蝫蝬蝭蝮蝯蝰蝱蝲蝳蝴蝶蝷蝸蝹蝺蝻蝼蝽蝾蝿螁螂螃螄螅螇螈螋融螎螏螐螑螒螓螔螖螗螘螚螛螜螝螞螟螠螡螢螣螤螥螨螪螫螬螭螮螯螰螱螳螴螵螶螷螸螹螺螻螼螽螾螿蟀蟁蟂蟃蟄蟅蟆蟇蟈蟉蟊蟋蟌蟎蟏蟑蟒蟓蟔蟗蟘蟙蟛蟜蟞蟟蟠蟡蟢蟣蟤蟥蟦蟧蟨蟪蟫蟬蟭蟮蟯蟲蟳蟴蟶蟷蟹蟺蟻蟼蟾蟿蠀蠁蠂蠃蠄蠅蠆蠈蠉蠊蠋蠌蠍蠏蠐蠑蠓蠔蠕蠖蠗蠘蠙蠛蠜蠝蠟蠠蠡蠢蠣蠤蠥蠦蠨蠩蠪蠫蠬蠭蠮蠯蠰蠱蠲蠳蠵蠶蠷蠸蠹蠻蠼蠽蠾蠿血衁衂衃衄衅衆衇衈衊衋行衍衎衒術衔衕衖街衙衚衛衜衝衞衠衡衢衣补衧表衩衪衫衬衭衮衯衰衱衲衵衶衷衹衺衽衾衿袀袁袂袃袄袅袆袈袉袋袌袍袑袒袓袕袖袗袘袙袚袛袜袞袟袠袡袢袤袧袨袪被袬袭袯袰袱袲袴袵袶袷袸袹袺袼袽袾袿裀裁裂裃装裆裈裉裊裋裍裎裏裒裔裕裖裗裘裙裚裛補裝裞裟裠裡裢裣裤裥裧裨裬裮裯裰裱裲裳裴裵裶裷裸裹裺裻裼製裾裿褀褁褂褄褅褆複褊褋褌褍褎褐褑褒褓褔褕褖褗褘褙褚褛褞褟褡褢褥褦褧褩褪褫褭褮褯褰褱褲褳褴褵褶褷褸褻褼褽褾襁襂襃襄襆襇襋襌襏襐襒襓襕襖襗襘襚襛襜襝襞襟襠襡襢襣襤襦襩襪襫襬襭襮襯襲襳襴襶襷襹襺襻襾西要覂覃覅覆覇覈見覌覎規覓覕視覗覘覚覛覜覝覡覢覣覤覦覧覩親覬覭覮覯覲観覶覷覺覽覿觀见观觃规觅视觇览觉觊觋觌觎觏觐觑角觓觔觕觖觙觚觜觝觞觟觠觡觢解觤觥触觩觫觬觭觯觰觱觲觳觴觶觷觸觺觻觼觾觿言訁訂訃訄訇計訊訌討訏訐訑訒訓訔訕訖託記訚訛訝訞訟訢訣訥訧訩訪訬設訰許訳訴訶訹診註証訾訿詀詁詄詅詆詈詍詎詐詒詔評詖詗詘詙詛詝詞詟詠詡詢詣詥試詧詩詪詫詬詭詮詰話該詳詵詶詷詹詻詼詿誂誃誄誅誆誇誉誊誋誌認誑誒誓誕誖誘誙誚語誠誡誣誤誥誦誨說誫説読誰課誴誶誸誹誻誼誾調諀諂諄諅諆談諈諉請諌諍諏諐諑諒諓諔諕論諗諚諛諜諝諞諟諠諡諢諤諦諧諪諫諭諮諰諱諲諳諴諵諶諷諸諺諼諾謀謁謂謄謅謇謈謊謋謌謍謎謏謐謑謒謔謕謖謗謘謙謚講謜謝謞謠謡謢謣謤謥謦謧謨謩謪謫謬謯謰謱謳謵謷謹謼謾譀譁譅譆譇譈證譊譋譌譎譏譐譑譓譔譕譖識譙譚譛譜譝譞譟譠譣譥警譨譪譫譬譭譯議譲譳譴護譸譹譺譻譽譾譿讀讂讃讄讆變讋讌讎讐讒讓讔讕讖讘讙讚讜讞讟计订讣认讥讦讧讨让讪讫训议讯记讱讲讳讴讵讶讷许讹论讻讼讽设访诀证诂诃评诅识诇诈诉诊诋诌词诎诏诐译诒诓诔试诖诗诘诙诚诛诜话诞诟诠诡询诣诤该详诧诨诩诫诬语诮误诰诱诲诳说诵请诸诹诺读诼诽课诿谀谁谂调谄谅谆谇谈谊谋谌谍谎谏谐谑谒谓谔谕谖谗谙谚谛谜谝谞谟谠谡谢谣谤谥谦谧谨谩谪谫谬谭谮谯谰谱谲谳谴谵谶谷谹谻谼谽谾谿豁豃豅豆豇豈豉豊豋豌豍豎豏豐豔豕豗豚豜豝豟象豢豤豥豦豨豪豫豬豭豮豯豰豱豲豳豵豶豷豸豹豺豻豽豿貀貁貂貄貅貆貉貊貌貍貏貐貑貒貓貔貕貗貘貙貛貜貝貞負財貢貣貤貧貨販貪貫責貭貮貯貰貲貳貴貶買貸貺費貼貽貾貿賀賁賂賃賄賅資賈賉賊賌賎賑賒賓賔賕賙賚賛賜賞賟賠賡賢賣賤賥賦賧賨質賫賬賭賮賴賵賸賹賺賻購賽賾贄贅贆贇贈贊贋贌贍贏贐贓贔贕贖贗贙贛贜贝贞负贡财责贤败账货质贩贪贫贬购贮贯贰贱贲贳贴贵贶贷贸费贺贻贼贽贾贿赀赁赂赃资赅赆赇赈赉赊赋赌赍赎赏赐赑赒赓赔赕赖赗赘赙赚赛赜赝赞赟赠赡赢赣赤赦赧赨赩赪赫赬赭赮走赳赴赵赶起赸趀趁趂趄超越趋趌趍趎趏趐趑趓趔趕趖趙趜趟趠趡趣趥趧趨趪趫趬趭趮趯趱趲足趴趵趷趸趹趺趼趾趿跂跃跄跅跆跇跈跋跌跍跎跏跐跑跓跕跖跗跘跙跚跛跜距跞跟跠跡跢跣跤跥跦跧跨跩跪跫跬跮路跰跱跲跳跴践跶跷跸跹跺跻跼跽跾跿踂踃踅踆踇踉踊踌踍踎踏踐踑踒踓踔踕踖踘踙踛踝踞踟踠踡踢踣踤踥踦踧踩踪踫踬踮踯踰踱踳踴踵踶踸踹踺踼踽踾蹀蹁蹂蹄蹅蹇蹈蹉蹊蹋蹌蹍蹎蹏蹐蹑蹒蹓蹔蹕蹖蹗蹙蹚蹛蹜蹝蹞蹟蹠蹡蹢蹣蹤蹥蹦蹧蹩蹪蹬蹭蹯蹰蹲蹳蹴蹵蹶蹸蹺蹻蹼蹽蹾蹿躁躂躄躅躆躇躈躉躊躋躍躏躐躑躒躓躔躕躖躗躙躚躜躝躞躟躠躡躣躤躥躦躨躩躪身躬躭躯躰躲躳躶躺躽軀車軋軌軍軎軏軑軒軓軔軘軛軝軞軟転軤軥軦軧軨軫軬軮軯軱軲軵軶軷軸軹軺軻軼軽軾軿輀輁輂較輄輅輆輇輈載輊輋輌輐輑輒輓輔輕輖輗輘輚輛輜輝輞輟輠輣輤輥輦輩輪輬輭輮輯輲輳輴輵輶輷輸輹輻輾輿轀轂轃轄轅轆轇轈轉轍轎轏轐轑轒轓轔轕轖轗轘轙轚轛轝轞轟轠轡轢轤车轧轨轩轪轫转轭轮软轰轱轲轳轴轵轶轷轸轹轺轻轼载轾轿辀辁辂较辄辅辆辇辈辉辊辋辌辍辎辏辐辑辒输辔辕辖辗辘辙辚辛辜辞辟辠辢辣辥辦辧辨辩辫辭辮辯辰辱農辴辵边辺辻込辽达辿迁迂迄迅迆过迈迉迋迍迎运近迒迓返迕迖还这进远违连迟迡迢迣迤迥迦迨迩迪迫迭迮述迳迴迵迷迸迹迺迻追迾迿退送适逃逄逅逆逈选逊逋逌逍逎透逐逑递逓途逕逖逗這通逛逜逝逞速造逡逢連逤逦逨逩逭逮逯週進逴逵逶逸逹逺逻逼逾逿遁遂遄遅遆遇遊運遍過遏遐遑遒道達違遗遘遙遛遜遝遞遠遡遢遣遥遨適遭遮遯遰遲遳遴遵遶遷選遹遺遻遼遽遾避邀邁邂邃還邅邆邇邈邉邊邋邍邏邐邑邓邔邕邗邘邙邛邝邞邟邠邡邢那邥邦邧邨邪邬邮邯邰邱邲邳邴邵邶邸邹邺邻邽邾邿郁郃郄郅郇郈郊郋郎郏郐郑郓郔郕郖郗郘郙郚郛郜郝郞郟郠郡郢郣郤郥郦郧部郪郫郭郯郰郱郲郳郴郵郷郸郹郺郻郼都郾郿鄀鄁鄂鄃鄄鄅鄆鄇鄈鄉鄋鄌鄍鄎鄏鄐鄑鄒鄔鄖鄗鄘鄙鄚鄛鄜鄝鄞鄟鄠鄡鄢鄣鄤鄦鄧鄨鄩鄪鄫鄬鄭鄮鄯鄰鄱鄲鄳鄴鄵鄶鄸鄹鄺鄻鄾酀酁酂酃酄酅酆酇酈酉酊酋酌配酎酏酐酒酓酔酖酗酘酚酝酞酟酡酢酣酤酥酦酧酨酩酪酬酮酯酰酱酲酳酴酵酶酷酸酹酺酼酽酾酿醁醂醃醅醆醇醉醊醋醌醍醐醑醒醓醕醖醗醙醚醛醜醞醟醢醣醤醥醧醨醪醫醬醭醮醯醰醱醲醳醴醵醷醸醹醺醻醼醽醾醿釀釁釂釃釅釆采釈釉释釋里重野量釐金釓釔釕釗釘釙釚釜針釢釣釤釦釧釩釪釬釭釱釳釴釵釷釸釹釺釽釿鈀鈁鈃鈄鈆鈇鈈鈉鈊鈌鈍鈎鈏鈐鈑鈒鈔鈕鈖鈚鈞鈢鈣鈥鈦鈧鈫鈮鈰鈳鈴鈶鈷鈸鈹鈺鈽鈾鈿鉀鉄鉅鉆鉇鉈鉉鉊鉋鉌鉍鉏鉑鉒鉓鉕鉗鉚鉛鉞鉟鉠鉢鉤鉥鉦鉧鉬鉭鉮鉯鉱鉲鉴鉶鉷鉸鉹鉺鉻鉼鉾鉿銀銂銃銅銈銋銍銎銑銓銔銕銖銗銘銚銛銜銠銣銤銥銦銨銩銪銫銬銭銮銱銲銳銶銷銹銻銼銾鋀鋁鋂鋃鋄鋅鋆鋇鋈鋊鋋鋌鋏鋐鋒鋕鋗鋘鋙鋝鋟鋡鋤鋥鋦鋨鋩鋪鋭鋮鋯鋰鋱鋲鋳鋶鋸鋹鋺鋼錀錁錄錆錈錍錎錏錐錒錕錘錙錚錛錞錟錠錡錢錣錦錧錨錫錬錭錮錯録錳錵錶錸錻錼錾鍀鍁鍆鍇鍈鍉鍊鍋鍌鍍鍏鍐鍑鍔鍖鍘鍚鍛鍜鍝鍟鍠鍤鍥鍩鍪鍬鍭鍮鍰鍱鍳鍵鍶鍷鍺鍼鍾鎂鎃鎅鎉鎊鎌鎎鎏鎒鎓鎔鎖鎗鎘鎚鎛鎝鎞鎡鎢鎣鎦鎧鎩鎪鎬鎭鎮鎯鎰鎳鎵鎷鎸鎹鎻鏂鏃鏇鏈鏊鏌鏍鏏鏐鏑鏓鏖鏗鏘鏙鏚鏜鏝鏞鏟鏡鏢鏤鏦鏨鏬鏰鏳鏴鏵鏷鏹鏺鏻鏽鏾鐀鐃鐇鐈鐋鐍鐎鐏鐐鐓鐔鐕鐖鐘鐙鐚鐝鐠鐡鐤鐦鐧鐨鐫鐬鐭鐮鐯鐲鐳鐵鐶鐸鐺鐻鐽鐾鐿鑀鑁鑂鑄鑅鑊鑋鑌鑐鑑鑒鑔鑕鑚鑛鑞鑠鑢鑣鑤鑨鑪鑫鑭鑮鑯鑰鑱鑲鑴鑵鑷鑹鑼鑽鑾鑿钀钁钂钃钆钇针钉钊钋钌钍钎钏钐钒钓钔钕钖钗钘钙钚钛钜钝钞钟钠钡钢钣钤钥钦钧钨钩钪钫钬钭钮钯钰钱钲钳钴钵钷钹钺钻钼钽钾钿铀铁铂铃铄铅铆铈铉铊铋铌铍铎铏铐铑铒铕铖铗铘铙铚铛铜铝铞铟铠铡铢铣铤铥铧铨铩铪铫铬铭铮铯铰铱铲铳铴铵银铷铸铹铺铻铼铽链铿销锁锂锃锄锅锆锇锈锉锊锋锌锍锎锏锐锑锒锓锔锕锖锗锘错锚锛锜锝锞锟锡锢锣锤锥锦锧锨锩锪锫锬锭键锯锰锱锲锳锴锵锶锷锸锹锺锻锼锽锾锿镀镁镂镃镄镅镆镇镈镉镊镋镌镍镎镏镐镑镒镓镔镕镖镗镘镚镛镜镝镞镠镡镢镣镤镥镦镧镨镩镪镫镬镭镮镯镰镱镲镳镴镵镶長镺镻镼镽长門閂閃閆閇閈閉開閌閍閎閏閑閒間閔閘閜閞閟閡関閣閤閥閦閧閨閩閫閬閭閰閱閲閵閶閷閹閺閻閼閽閾閿闀闃闅闆闇闈闉闊闋闌闍闐闑闒闓闔闕闖闘闚闛關闞闟闠闡闢闤闥门闩闪闫闭问闯闰闱闲闳间闵闶闷闸闹闺闻闼闽闾闿阀阁阂阃阄阅阆阇阈阉阊阋阌阍阎阏阐阑阒阔阕阖阗阘阙阚阜阝阞队阠阡阢阤阨阪阬阭阮阯阰阱防阳阴阵阶阸阹阺阻阼阽阿陀陁陂附际陆陇陈陉陊陋陌降陎陏限陑陓陔陕陗陘陛陜陝陞陟陡院陣除陥陧陨险陪陫陬陭陰陲陳陴陵陶陷陸険陻陼陽陾隃隄隅隆隇隈隉隊隋隍階随隐隑隒隓隔隕隗隘隙際障隞隠隣隤隧隨隩險隮隰隱隲隳隴隶隷隸隹隺隻隼隽难隿雀雁雂雄雅集雇雈雉雊雋雌雍雎雏雑雒雓雔雕雖雗雘雙雚雛雜雝雞雟雠離難雨雩雪雫雯雰雱雲雳零雷雹雺電雽雾雿需霁霂霄霅霆震霈霉霊霋霍霎霏霐霑霒霓霖霙霜霞霠霢霣霤霧霨霩霪霫霬霭霮霰露霵霶霸霹霽霾霿靁靂靃靄靆靈靉靑青靓靖静靚靛靜非靠靡面靥靦靨革靪靫靬靭靮靰靱靲靳靴靶靷靸靺靻靼靽靾靿鞀鞁鞂鞃鞄鞅鞆鞈鞊鞋鞌鞍鞎鞏鞑鞒鞔鞗鞘鞙鞚鞜鞞鞠鞡鞣鞤鞥鞦鞧鞨鞪鞫鞬鞭鞮鞯鞲鞳鞴鞵鞶鞷鞹鞻鞽鞾鞿韁韂韃韄韅韆韇韉韋韌韍韎韏韐韑韓韔韕韗韙韜韝韞韟韠韡韣韥韦韧韨韩韪韫韬韭韮韰韱音韵韶韹韺韻韽韾響頁頂頃頄項順頇須頊頌頍頎頏預頑頒頓頔頖頗領頚頜頝頞頠頡頤頦頨頩頫頬頭頯頰頲頴頵頷頸頹頻頼頽顁顄顅顆顇顈顉顊顋題額顎顏顐顑顒顓顔顕顗願顙顛顜顝類顟顠顢顣顤顥顦顧顩顪顫顯顰顱顲顳顴页顶顷顸项顺须顼顽顾顿颀颁颂颃预颅领颇颈颉颊颋颌颍颎颏颐频颓颔颖颗题颙颚颛颜额颞颟颠颡颢颤颥颦颧風颬颭颮颯颱颲颳颶颸颺颻颼颽颾颿飁飂飄飆飈飉飋飌风飏飐飑飒飓飔飕飗飘飙飛飜飞食飡飢飣飥飧飨飩飪飫飬飭飮飯飱飲飴飶飺飼飽飾飿餀餂餃餄餅餈餉養餌餍餎餏餐餑餒餓餔餕餖餗餘餚餛餜餝餞餟餠餡餤餥餧館餪餫餬餭餮餯餰餱餲餳餵餷餹餺餻餼餽餾餿饁饂饃饅饇饈饉饊饋饌饍饎饐饑饒饔饕饗饘饙饛饜饞饟饡饢饥饧饨饩饪饫饬饭饮饯饰饱饲饳饴饵饶饷饸饹饺饻饼饽饿馁馃馄馅馆馇馈馉馊馋馌馍馏馐馑馒馓馔馕首馗馘香馛馝馞馣馥馦馧馨馬馭馮馯馰馱馲馳馴馵馹馺馻馼馽駁駂駃駄駅駆駈駉駍駎駏駐駑駒駓駔駕駖駘駙駛駜駝駟駡駢駣駤駥駩駪駬駭駮駰駱駴駶駷駸駹駺駻駼駽駾駿騁騂騃騄騅騇騉騊騋騌騍騎騏騑騒験騕騖騙騚騛騜騝騞騠騢騣騤騥騧騨騩騪騫騬騭騮騰騱騲騴騵騶騷騸騹騺騽騾驀驁驂驃驄驅驆驈驉驊驌驍驎驏驒驓驔驕驖驗驚驛驞驟驠驢驤驥驦驨驩驪驫马驭驮驯驰驱驲驳驴驵驶驷驸驹驺驻驼驽驾驿骀骁骂骃骄骅骆骇骈骉骊骋验骍骎骏骐骑骒骓骕骖骗骘骙骚骛骜骝骞骟骠骡骢骣骤骥骦骧骨骫骭骯骰骱骴骶骷骸骹骺骼骽骾骿髀髁髂髃髄髅髆髇髈髊髋髌髍髎髏髐髑髒髓體髕髖高髙髟髡髢髣髥髦髧髪髫髬髭髮髯髲髳髴髶髷髹髺髻髼髽髾鬀鬁鬃鬄鬅鬆鬈鬉鬊鬋鬌鬍鬎鬏鬐鬒鬓鬕鬖鬗鬘鬙鬚鬞鬟鬠鬢鬣鬤鬥鬧鬨鬩鬪鬫鬬鬮鬯鬱鬲鬳鬵鬶鬷鬹鬺鬻鬼鬾鬿魁魂魃魄魅魆魇魈魉魊魋魌魍魎魏魑魔魕魖魘魚魛魞魟魠魡魣魦魧魨魩魮魯魱魴魵魶魷魻魼魾鮀鮁鮂鮃鮄鮅鮆鮇鮊鮋鮍鮎鮐鮑鮒鮓鮗鮚鮛鮜鮝鮞鮟鮠鮡鮣鮤鮥鮦鮨鮪鮫鮭鮮鮰鮵鮸鮹鮻鮿鯀鯁鯃鯆鯇鯈鯉鯊鯏鯒鯓鯔鯕鯖鯗鯙鯛鯝鯞鯠鯡鯢鯤鯦鯧鯨鯪鯫鯬鯭鯮鯰鯱鯶鯷鯸鯻鯽鯿鰂鰃鰅鰆鰈鰉鰋鰌鰍鰏鰐鰒鰓鰕鰗鰛鰜鰝鰟鰣鰤鰥鰧鰨鰩鰫鰬鰭鰮鰯鰰鰱鰲鰳鰴鰶鰷鰹鰺鰻鰼鰾鰿鱀鱁鱂鱄鱅鱆鱇鱈鱉鱊鱋鱍鱐鱒鱓鱔鱕鱖鱗鱘鱚鱝鱞鱟鱠鱢鱣鱥鱦鱧鱨鱬鱭鱮鱰鱲鱳鱴鱵鱷鱸鱹鱺鱻鱼鱽鱾鱿鲀鲁鲂鲃鲅鲆鲇鲈鲉鲊鲋鲌鲍鲎鲏鲐鲑鲒鲔鲕鲖鲗鲘鲙鲚鲛鲜鲝鲞鲟鲠鲡鲢鲣鲤鲥鲦鲧鲨鲩鲪鲫鲬鲭鲮鲯鲰鲱鲲鲳鲴鲵鲷鲸鲹鲺鲻鲼鲽鲾鲿鳀鳁鳂鳃鳄鳅鳇鳈鳉鳊鳌鳍鳎鳏鳐鳑鳒鳓鳔鳕鳖鳗鳘鳙鳚鳛鳜鳝鳞鳟鳠鳡鳢鳣鳤鳥鳦鳧鳩鳪鳬鳭鳯鳰鳱鳲鳳鳴鳵鳶鳷鳹鳺鳻鳼鳽鳾鳿鴀鴃鴅鴆鴇鴈鴉鴎鴐鴒鴓鴔鴕鴗鴘鴛鴝鴞鴟鴠鴡鴢鴣鴥鴦鴨鴩鴫鴮鴯鴰鴱鴳鴴鴶鴷鴸鴻鴽鴾鴿鵀鵁鵂鵃鵄鵅鵊鵋鵌鵎鵏鵐鵑鵒鵓鵔鵖鵗鵙鵛鵜鵝鵞鵟鵠鵡鵧鵩鵪鵫鵬鵮鵯鵰鵱鵲鵳鵴鵵鵷鵸鵹鵺鵻鵼鵽鵾鶀鶁鶂鶄鶅鶆鶇鶈鶉鶊鶋鶌鶏鶐鶒鶓鶔鶖鶗鶘鶙鶚鶛鶜鶝鶞鶟鶠鶡鶢鶣鶤鶥鶨鶩鶪鶬鶭鶯鶱鶲鶴鶵鶶鶷鶹鶺鶻鶼鶾鶿鷀鷁鷂鷃鷄鷅鷇鷈鷉鷊鷋鷎鷏鷐鷑鷒鷓鷕鷖鷗鷘鷙鷚鷛鷜鷞鷟鷡鷢鷣鷤鷥鷦鷩鷫鷮鷯鷰鷲鷳鷵鷶鷷鷸鷹鷺鷻鷽鷾鷿鸀鸁鸂鸃鸄鸅鸆鸇鸉鸊鸋鸌鸍鸏鸐鸑鸒鸓鸔鸕鸗鸘鸙鸚鸛鸜鸝鸞鸟鸠鸡鸢鸣鸤鸥鸦鸧鸨鸩鸪鸫鸬鸭鸮鸯鸰鸱鸲鸳鸵鸶鸷鸸鸹鸺鸻鸼鸽鸾鸿鹀鹁鹂鹃鹄鹅鹆鹇鹈鹉鹊鹋鹌鹍鹎鹏鹐鹑鹒鹔鹕鹖鹗鹘鹙鹚鹛鹜鹝鹞鹟鹠鹡鹢鹣鹤鹦鹧鹨鹩鹪鹫鹬鹭鹮鹯鹰鹱鹲鹳鹴鹵鹸鹹鹺鹻鹼鹽鹾鹿麀麁麂麃麅麆麇麈麉麊麋麌麎麐麑麒麓麔麖麗麙麚麛麜麝麞麟麠麡麤麥麦麧麩麪麮麯麰麴麵麶麷麸麹麺麻麼麽麾麿黀黁黂黃黄黇黈黉黌黍黎黏黐黑黒黓黔黕黖默黙黚黛黜黝點黟黠黡黢黤黥黧黨黩黪黫黭黮黯黰黲黳黴黵黶黷黹黺黻黼黽黾黿鼀鼁鼂鼃鼆鼇鼈鼉鼊鼋鼍鼎鼏鼐鼒鼓鼕鼖鼗鼘鼙鼛鼜鼠鼢鼣鼤鼥鼨鼩鼪鼫鼬鼭鼮鼯鼰鼱鼲鼳鼴鼵鼶鼷鼸鼹鼻鼽鼾齀齁齂齃齇齉齊齋齌齍齎齏齐齑齒齔齕齖齗齘齙齛齜齝齞齟齠齡齢齣齤齥齦齧齩齪齫齬齮齯齰齱齲齴齵齶齷齸齹齺齻齾齿龀龁龂龃龄龅龆龇龈龉龊龋龌龍龎龐龑龒龔龕龗龙龚龛龜龝龟龠龢龤鿍鿎鿏郎凉︽︾﹏﹥﹪！＃％＆（）＋，－．／０１２３４５６７８９：；＝？＠ＡＢＣＤＥＦＧＨＩＪＫＬＭＮＯＰＱＲＳＴＵＶＷＸＹＺａｂｃｄｅｆｇｈｉｊｋｌｍｎｏｐｑｒｓｔｕｖｗｘｙｚ｜～･ｦｧｨｩｪｫｬｭｮｯｱｲｳｴｵｶｷｸｹｺｻｼｽｾｿﾀﾁﾂﾃﾄﾅﾆﾇﾈﾉﾊﾋﾌﾍﾎﾏﾐﾑﾒﾓﾔﾕﾖﾗﾘﾙﾚﾛﾜﾝﾞﾟ￥𠅤𠙶𠳐𠴰𡎚𡐓𡑍𡚁𣇉𣗋𣲗𣲘𣸣𤋮𤧛𤩽𤫉𤺥𥔲𥕢𥖨𥻗𦈡𦍑𦒍𦙶𦝼𦭜𦰡𧿹𨐈𨙸𨚕𨟠𨭉𨱇𨱏𨱑𨱔𨺙𩓐𩽾𩾃𩾌𪟝𪣻𪤗𪨰𪨶𪩘𪾢𫄧𫄨𫄷𫄸𫇭𫌀𫍣𫍯𫍲𫍽𫐄𫐐𫐓𫑡𫓧𫓯𫓶𫓹𫔍𫔎𫔶𫖮𫖯𫖳𫗧𫗴𫘜𫘝𫘦𫘧𫘨𫘪𫘬𫚕𫚖𫚭𫛭𫞩𫟅𫟦𫟹𫟼𫠆𫠊𫠜𫢸𫫇𫭟𫭢𫭼𫮃𫰛𫵷𫶇𫷷𫸩𬀩𬀪𬂩𬃊𬇕𬇙𬇹𬉼𬊈𬊤𬌗𬍛𬍡𬍤𬒈𬒔𬒗𬕂𬘓𬘘𬘡𬘩𬘫𬘬𬘭𬘯𬙂𬙊𬙋𬜬𬜯𬞟𬟁𬟽𬣙𬣞𬣡𬣳𬤇𬤊𬤝𬨂𬨎𬩽𬪩𬬩𬬭𬬮𬬱𬬸𬬹𬬻𬬿𬭁𬭊𬭎𬭚𬭛𬭤𬭩𬭬𬭯𬭳𬭶𬭸𬭼𬮱𬮿𬯀𬯎𬱖𬱟𬳵𬳶𬳽𬳿𬴂𬴃𬴊𬶋𬶍𬶏𬶐𬶟𬶠𬶨𬶭𬶮𬷕𬸘𬸚𬸣𬸦𬸪𬹼𬺈𬺓🀄🃏🅰🅱🅾🅿🆎🆑🆒🆓🆔🆕🆖🆗🆘🆙🆚🇨🇩🇪🇫🇬🇮🇯🇰🇷🇺🈁🌀🌁🌂🌃🌄🌅🌆🌇🌈🌉🌊🌋🌌🌍🌎🌏🌐🌑🌒🌓🌔🌕🌖🌗🌘🌙🌚🌛🌜🌝🌞🌟🌠🌰🌱🌲🌳🌴🌵🌷🌸🌹🌺🌻🌼🌽🌾🌿🍀🍁🍂🍃🍄🍅🍆🍇🍈🍉🍊🍋🍌🍍🍎🍏🍐🍑🍒🍓🍔🍕🍖🍗🍘🍙🍚🍛🍜🍝🍞🍟🍠🍡🍢🍣🍤🍥🍦🍧🍨🍩🍪🍫🍬🍭🍮🍯🍰🍱🍲🍳🍴🍵🍶🍷🍸🍹🍺🍻🍼🎀🎁🎂🎃🎄🎅🎆🎇🎈🎉🎊🎋🎌🎍🎎🎏🎐🎑🎒🎓🎠🎡🎢🎣🎤🎥🎦🎧🎨🎩🎪🎫🎬🎭🎮🎯🎰🎱🎲🎳🎴🎵🎶🎷🎸🎹🎺🎻🎼🎽🎾🎿🏀🏁🏂🏃🏄🏆🏇🏈🏉🏊🏠🏡🏢🏣🏤🏥🏦🏧🏨🏩🏪🏫🏬🏭🏮🏯🏰🐀🐁🐂🐃🐄🐅🐆🐇🐈🐉🐊🐋🐌🐍🐎🐏🐐🐑🐒🐓🐔🐕🐖🐗🐘🐙🐚🐛🐜🐝🐞🐟🐠🐡🐢🐣🐤🐥🐦🐧🐨🐩🐪🐫🐬🐭🐮🐯🐰🐱🐲🐳🐴🐵🐶🐷🐸🐹🐺🐻🐼🐽🐾👀👂👃👄👅👆👇👈👉👊👋👌👍👎👏👐👑👒👓👔👕👖👗👘👙👚👛👜👝👞👟👠👡👢👣👤👥👦👧👨👩👪👫👬👭👮👯👰👱👲👳👴👵👶👷👸👹👺👻👼👽👾👿💀💁💂💃💄💅💆💇💈💉💊💋💌💍💎💏💐💑💒💓💔💕💖💗💘💙💚💛💜💝💞💟💠💡💢💣💤💥💦💧💨💩💪💫💬💭💮💯💰💲💳💴💵💶💷💸💹💺💻💼💽💾💿📀📁📂📃📄📅📆📇📈📉📊📋📌📍📎📏📐📑📒📓📔📕📖📗📘📙📚📛📜📝📞📟📠📡📢📣📤📥📦📧📨📩📪📫📬📭📮📯📰📱📲📳📴📵📶📷📹📺📻📼🔀🔁🔂🔃🔄🔅🔆🔇🔉🔊🔋🔌🔍🔎🔏🔐🔑🔒🔓🔔🔕🔖🔗🔘🔙🔚🔛🔜🔝🔞🔟🔠🔡🔢🔣🔤🔥🔦🔧🔨🔩🔪🔫🔬🔭🔮🔯🔰🔱🔲🔳🔴🔵🔶🔷🔸🔹🔺🔻🔼🔽🕐🕑🕒🕓🕔🕕🕖🕗🕘🕙🕚🕛🕜🕝🕞🕟🕠🕡🕢🕣🕤🕥🕦🕧🗻🗼🗽🗾🗿😀😁😂😃😄😅😆😇😈😉😊😋😌😍😎😏😐😑😒😓😔😕😖😗😘😙😚😛😜😝😞😟😠😡😢😣😤😥😦😧😨😩😪😫😬😭😮😯😰😱😲😳😴😵😶😷😸😹😺😻😼😽😾😿🙀🙅🙆🙇🙈🙉🙊🙋🙌🙍🙎🙏🚀🚁🚂🚃🚄🚅🚆🚇🚈🚉🚊🚌🚍🚎🚏🚐🚑🚒🚓🚔🚕🚖🚗🚘🚙🚚🚛🚜🚝🚞🚟🚠🚡🚢🚣🚤🚥🚦🚧🚨🚩🚪🚫🚬🚭🚮🚯🚰🚱🚲🚳🚴🚵🚶🚷🚸🚹🚺🚻🚼🚽🚾🚿🛀🛁🛂🛃🛄🛅 ';

From 742986e0e3223b5bae1b91def63ef2326e40dc34 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 00:22:29 +0200
Subject: [PATCH 02/29] [RNE Rewrite] fix(ocr): valid DBNet reading-order
 comparator + dewarp off by default

- ocr_ops.cpp: quantise DBNet quad y-coordinates into fixed row bands before
  sorting. The previous `|dy| > 10` comparator was not a strict-weak ordering
  (intransitive), which aborts under libc++ hardening.
- document demo: default dewarp OFF. UVDoc dewarp only helps photographed,
  physically-warped pages; on flat images it distorts otherwise-clean text.
  Updated the screen copy accordingly.
---
 apps/computer-vision/app/document/index.tsx           | 10 ++++++----
 .../cpp/extensions/cv/ocr_ops.cpp                     | 11 +++++++++--
 2 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/apps/computer-vision/app/document/index.tsx b/apps/computer-vision/app/document/index.tsx
index 3aaf1c070a..857ea7362b 100644
--- a/apps/computer-vision/app/document/index.tsx
+++ b/apps/computer-vision/app/document/index.tsx
@@ -84,7 +84,9 @@ function DocumentContent() {
   const [layoutOn, setLayoutOn] = useState(true);
   const [supportingOn, setSupportingOn] = useState(true);
   const [orientation, setOrientation] = useState(true);
-  const [dewarp, setDewarp] = useState(true);
+  // Off by default: dewarp (UVDoc) corrects photographed, physically-warped pages;
+  // on a flat screenshot it has nothing to fix and visibly distorts clean text.
+  const [dewarp, setDewarp] = useState(false);
   const [imageUri, setImageUri] = useState<string | null>(null);
   const [isProcessing, setIsProcessing] = useState(false);
   const [blocks, setBlocks] = useState<DocBlock[]>([]);
@@ -177,9 +179,9 @@ function DocumentContent() {
       contentContainerStyle={commonStyles.contentContainer}
     >
       <Text style={commonStyles.description}>
-        Full document pipeline: layout → OCR grouped into reading-ordered blocks, with
-        orientation/dewarp and table-structure recognition. PaddleOCR is always on; the rest are on
-        by default — toggle any off (toggling reloads the models).
+        Full document pipeline: layout → OCR grouped into reading-ordered blocks, with orientation,
+        table-structure recognition and (optional) dewarp. PaddleOCR is always on; dewarp is off by
+        default — it only helps photographed, warped pages (toggling reloads the models).
       </Text>
 
       <ModelPicker
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 19ab24e6ce..1d3fe6c685 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -491,9 +491,16 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, floa
         quads.push_back(q);
     }
 
+    // Reading order: top -> bottom by ~row, then left -> right. Quantise y into
+    // row bands first so the comparator is a valid strict-weak ordering — a raw
+    // `|dy| > threshold` test is intransitive (a~b, b~c, but a<c) and aborts under
+    // libc++ hardening.
+    constexpr float kRowBand = 10.0f;
     std::sort(quads.begin(), quads.end(), [](const Quad &a, const Quad &b) {
-        if (std::abs(a.pts[0].y - b.pts[0].y) > 10.0f) {
-            return a.pts[0].y < b.pts[0].y;
+        const int rowA = static_cast<int>(std::floor(a.pts[0].y / kRowBand));
+        const int rowB = static_cast<int>(std::floor(b.pts[0].y / kRowBand));
+        if (rowA != rowB) {
+            return rowA < rowB;
         }
         return a.pts[0].x < b.pts[0].x;
     });

From 1fea8bd9383b6dc03633adcb591d39bc2fa3221a Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 02:37:45 +0200
Subject: [PATCH 03/29] [RNE Rewrite] fix(ocr): vertical glyph-split,
 per-method unload, per-run orientation/dewarp
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Verified on-device (Android emulator, PaddleOCR/XNNPACK):

- Vertical OCR (ocr.ts): stacked columns were detected/placed correctly but
  read as garbage (a vertical "ANTIQUES" → " 1 "). Root cause: the DBNet
  detector emits one box per text region, not per glyph, fusing stacked letters
  into a few tall boxes; recognizeGlyphStrip warped each multi-letter box into a
  single recognizer cell (squashed → garbage), and the char-level re-detect path
  doesn't split for DBNet. Fix: add splitTallQuad() and split every glyph box
  into ~square single-letter cells (by height/width) before strip assembly.
  Now reads "ANTIQUES"/"PARKING" at 91-93% (both the column and tall-single
  paths). [Bug 2]

- Bucketed-OCR memory (model.cpp + core/model.ts + ocr.ts + documentOCR.ts):
  each detect_<S>/recognize_<W> method's planned-memory arena was cached for the
  model's lifetime, so memory grew unbounded as image/box sizes varied (worst on
  CoreML, one compiled graph per method). Ported main's unload-after-use via the
  ET API: expose Model.unloadMethod() (Module::unload_method) and free the bucket
  arenas after each top-level run (RunOCROptions.release, default true; the
  document orchestrator frees once per page). Measured: a 640→1280 two-image run
  holds ~693 MB native heap without unload vs ~341 MB with it. [Bug 3]

- Document orientation/dewarp (documentOCR.ts + document demo): were baked at
  createDocumentOCR time, so useModel never recreated the model on toggle and the
  switches did nothing. Made them per-run options on runDocumentOCR(input,
  {orientation, dewarp}) (mirroring OCR's vertical), defaulting to the config
  flags; toggles now take effect with no reload. [Bug 4 / in-flight]

Dewarp (Bug 1) needed no code change: on-device the gridSample [-1,1] backward-map
convention is correct (near-identity grid on flat pages, correctly flattens a
warped page); the mild flat-page distortion is UVDoc emitting a non-identity field
and is indistinguishable from a real warp by the grid alone, so default-OFF +
the per-run toggle is the right mitigation.
---
 apps/computer-vision/app/document/index.tsx   |  24 +--
 .../cpp/core/model.cpp                        |  34 +++++
 .../react-native-executorch/src/core/model.ts |  15 ++
 .../src/extensions/cv/tasks/documentOCR.ts    | 142 ++++++++++++------
 .../src/extensions/cv/tasks/ocr.ts            | 103 +++++++++++--
 5 files changed, 244 insertions(+), 74 deletions(-)

diff --git a/apps/computer-vision/app/document/index.tsx b/apps/computer-vision/app/document/index.tsx
index 857ea7362b..28b95fe50e 100644
--- a/apps/computer-vision/app/document/index.tsx
+++ b/apps/computer-vision/app/document/index.tsx
@@ -101,12 +101,12 @@ function DocumentContent() {
   const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
 
   // Hosted configs — `useDocumentOCR` downloads + caches each enabled model.
+  // orientation/dewarp are NOT baked here: they're passed per-run to
+  // `runDocumentOCR` below, so toggling them takes effect without a reload.
   const config = {
     ocr: models.ocr.PADDLE.PPOCRV6_SMALL[backend.key],
     ...(layoutOn ? { layout: models.layoutDetection.PP_DOCLAYOUT[backend.key] } : {}),
     ...(supportingOn ? { supporting: models.supporting.PP_SUPPORTING[backend.key] } : {}),
-    orientation,
-    dewarp,
   };
 
   const { isReady, downloadProgress, error: loadError, runDocumentOCR } = useDocumentOCR(config);
@@ -134,13 +134,16 @@ function DocumentContent() {
       const pixels = skiaImage.readPixels();
       if (!(pixels instanceof Uint8Array)) throw new Error('Expected Uint8Array from readPixels');
       const start = Date.now();
-      const out = await runDocumentOCR({
-        data: pixels,
-        width: skiaImage.width(),
-        height: skiaImage.height(),
-        format: 'rgba' as const,
-        layout: 'hwc' as const,
-      });
+      const out = await runDocumentOCR(
+        {
+          data: pixels,
+          width: skiaImage.width(),
+          height: skiaImage.height(),
+          format: 'rgba' as const,
+          layout: 'hwc' as const,
+        },
+        { orientation, dewarp }
+      );
       setWallMs(Date.now() - start);
       setBlocks(out.blocks as DocBlock[]);
       // Show the frame the boxes are relative to (orientation/dewarp may have
@@ -181,7 +184,8 @@ function DocumentContent() {
       <Text style={commonStyles.description}>
         Full document pipeline: layout → OCR grouped into reading-ordered blocks, with orientation,
         table-structure recognition and (optional) dewarp. PaddleOCR is always on; dewarp is off by
-        default — it only helps photographed, warped pages (toggling reloads the models).
+        default — it only helps photographed, warped pages. Orientation/dewarp are per-run, so
+        toggling them takes effect on the next run without reloading the models.
       </Text>
 
       <ModelPicker
diff --git a/packages/react-native-executorch/cpp/core/model.cpp b/packages/react-native-executorch/cpp/core/model.cpp
index ba50d9d12b..f4b11c4e4b 100644
--- a/packages/react-native-executorch/cpp/core/model.cpp
+++ b/packages/react-native-executorch/cpp/core/model.cpp
@@ -433,6 +433,39 @@ jsi::Value ModelHostObject::get(jsi::Runtime &rt, const jsi::PropNameID &name) {
         return jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, "execute"), 3, fnBody);
     }
 
+    if (nameStr == "unloadMethod") {
+        auto self = shared_from_this();
+        auto fnBody = [self](jsi::Runtime &rt, const jsi::Value & /*thisVal*/, const jsi::Value *args, size_t count) -> jsi::Value {
+            if (count != 1) {
+                throw jsi::JSError(rt, "unloadMethod: Usage: unloadMethod(methodName)");
+            }
+
+            if (!args[0].isString()) {
+                throw jsi::JSError(rt, "unloadMethod: Expected arg0 to be a string");
+            }
+
+            std::unique_lock<std::mutex> lock(self->mutex_, std::try_to_lock);
+            if (!lock.owns_lock()) {
+                throw jsi::JSError(rt, "unloadMethod: Model is currently in use");
+            }
+
+            if (!self->etModule_) {
+                throw jsi::JSError(rt, "unloadMethod: Model has been disposed");
+            }
+
+            // Free a single previously-executed method's planned-memory activation
+            // arena (and, on graph-compiling backends like CoreML, its compiled
+            // graph). The method transparently reloads on next execute. Returns
+            // whether a loaded method was actually freed (false = not loaded, a
+            // harmless no-op). Bounds memory when many distinct bucketed methods
+            // (detect_<S>/recognize_<W>) accumulate over a session.
+            auto methodName = args[0].asString(rt).utf8(rt);
+            bool unloaded = self->etModule_->unload_method(methodName);
+            return jsi::Value(unloaded);
+        };
+        return jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, "unloadMethod"), 1, fnBody);
+    }
+
     if (nameStr == "dispose") {
         auto self = shared_from_this();
         auto fnBody = [self](jsi::Runtime &rt, const jsi::Value & /*thisVal*/, const jsi::Value * /*args*/, size_t count) -> jsi::Value {
@@ -462,6 +495,7 @@ std::vector<facebook::jsi::PropNameID> ModelHostObject::getPropertyNames(jsi::Ru
     properties.push_back(jsi::PropNameID::forAscii(rt, "getMethodNames"));
     properties.push_back(jsi::PropNameID::forAscii(rt, "getMethodMeta"));
     properties.push_back(jsi::PropNameID::forAscii(rt, "execute"));
+    properties.push_back(jsi::PropNameID::forAscii(rt, "unloadMethod"));
     properties.push_back(jsi::PropNameID::forAscii(rt, "dispose"));
     return properties;
 }
diff --git a/packages/react-native-executorch/src/core/model.ts b/packages/react-native-executorch/src/core/model.ts
index e539afecdd..5b7274d6ec 100644
--- a/packages/react-native-executorch/src/core/model.ts
+++ b/packages/react-native-executorch/src/core/model.ts
@@ -119,6 +119,21 @@ export interface Model {
    */
   execute(methodName: string, inputs: ModelInput[], outputTensors: Tensor[]): ModelOutput[];
 
+  /**
+   * Unloads a single previously-executed method, freeing its memory-planned
+   * activation arena (and, on graph-compiling backends like CoreML, its
+   * compiled graph). The method transparently reloads on its next `execute`.
+   *
+   * Use this to bound native memory when many distinct methods are executed
+   * over a session — e.g. bucketed OCR, where each `detect_<S>`/`recognize_<W>`
+   * size that is ever run would otherwise stay resident for the model's
+   * lifetime.
+   * @param methodName The exported method to unload.
+   * @returns `true` if a loaded method was freed, `false` if it was not loaded
+   * (a harmless no-op).
+   */
+  unloadMethod(methodName: string): boolean;
+
   /**
    * Releases the native ExecuTorch model and frees all associated resources.
    *
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
index ad8e751d0d..ac2e5c2daa 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -55,16 +55,37 @@ export type DocumentResult<L> = {
  * Configuration for the document OCR orchestrator. Provides an OCR model, an
  * optional layout model (regions/blocks), and an optional supporting model
  * (orientation/dewarp pre-processing + table-structure recognition). The
- * `orientation`/`dewarp` flags gate those pre-processing passes.
+ * `orientation`/`dewarp` flags are *defaults* for the per-run options of the
+ * same name — supply them here to bias every run, or leave them off and pass
+ * them to `runDocumentOCR` per call (the supporting model is loaded either way).
  * @category Types
  */
 export type DocumentOCRModel<L> = {
   readonly ocr: OCRModel;
   readonly layout?: ObjectDetectorModel<'xyxy', L>;
   readonly supporting?: SupportingModel;
-  /** Detect + correct page orientation before OCR (needs `supporting`). */
+  /** Default for the per-run `orientation` option (needs `supporting`). */
   readonly orientation?: boolean;
-  /** Geometrically dewarp the page before OCR (needs `supporting`). */
+  /** Default for the per-run `dewarp` option (needs `supporting`). */
+  readonly dewarp?: boolean;
+};
+
+/**
+ * Per-run document options (passed to `runDocumentOCR`, not baked into the
+ * model — toggling them needs no reload). Each pre-processing pass still
+ * requires the supporting model to have been loaded (`config.supporting`).
+ * @category Types
+ */
+export type RunDocumentOCROptions = {
+  /**
+   * Detect + correct page orientation before OCR. No-op without a loaded
+   * supporting model. Defaults to the model's `config.orientation`.
+   */
+  readonly orientation?: boolean;
+  /**
+   * Geometrically dewarp the page before OCR. No-op without a loaded supporting
+   * model. Defaults to the model's `config.dewarp`.
+   */
   readonly dewarp?: boolean;
 };
 
@@ -133,16 +154,21 @@ export async function createDocumentOCR<L>(
   runtime?: WorkletRuntime
 ): Promise<{
   dispose: () => void;
-  runDocumentOCR: (input: ImageBuffer) => Promise<DocumentResult<L>>;
-  runDocumentOCRWorklet: (input: ImageBuffer) => DocumentResult<L>;
+  runDocumentOCR: (
+    input: ImageBuffer,
+    options?: RunDocumentOCROptions
+  ) => Promise<DocumentResult<L>>;
+  runDocumentOCRWorklet: (input: ImageBuffer, options?: RunDocumentOCROptions) => DocumentResult<L>;
 }> {
   const ocr = await createOCR(config.ocr, runtime);
   const layout = config.layout
     ? await createObjectDetector<'xyxy', L>(config.layout, runtime)
     : null;
   const supporting = config.supporting ? await createSupporting(config.supporting, runtime) : null;
-  const useOrientation = !!supporting && !!config.orientation;
-  const useDewarp = !!supporting && !!config.dewarp;
+  // Per-run orientation/dewarp default to the model's config flags; both are
+  // no-ops without a loaded supporting model.
+  const defaultOrientation = !!config.orientation;
+  const defaultDewarp = !!config.dewarp;
 
   const dispose = () => {
     ocr.dispose();
@@ -150,8 +176,13 @@ export async function createDocumentOCR<L>(
     supporting?.dispose();
   };
 
-  const runDocumentOCRWorklet = (input: ImageBuffer): DocumentResult<L> => {
+  const runDocumentOCRWorklet = (
+    input: ImageBuffer,
+    options?: RunDocumentOCROptions
+  ): DocumentResult<L> => {
     'worklet';
+    const useOrientation = !!supporting && (options?.orientation ?? defaultOrientation);
+    const useDewarp = !!supporting && (options?.dewarp ?? defaultDewarp);
     let img = input;
     if (useOrientation && supporting) {
       const rot = supporting.detectOrientationWorklet(img).rotationCW;
@@ -163,50 +194,63 @@ export async function createDocumentOCR<L>(
       img = supporting.dewarpWorklet(img);
     }
 
-    // Mode A — no layout: OCR the whole page into one block.
-    if (!layout) {
-      const detections = ocr.runOCRWorklet(img).detections;
-      const blocks = detections.length
-        ? [
-            makeBlock<L>(
-              'ungrouped',
-              boundingBoxOf(detections.flatMap((d) => d.quad as Point[])),
-              1,
-              detections,
-              false
-            ),
-          ]
-        : [];
-      return { blocks, regions: [], detections, image: img };
-    }
-
-    // Mode B — layout: OCR each text region's crop on its own (upscaled into the
-    // detector → far better recall than one whole-page pass), offsetting lines
-    // back to page coords. Tables also recognize structure + fill cells.
-    const regions = layout.detectObjectsWorklet(img);
-    const blocks: DocumentBlock<L>[] = [];
-    const detections: OCRDetection[] = [];
-    for (const region of regions) {
-      if (!isTextRegion(region.label)) {
-        continue;
-      }
-      const { xmin, ymin } = region.box;
-      const crop = cropImageBuffer(img, region.box);
-      const lines = ocr.runOCRWorklet(crop).detections.map((d) => offsetDetection(d, xmin, ymin));
-      const table = isTableLabel(region.label);
-      if (lines.length === 0 && !table) {
-        continue;
+    // OCR runs once per region here (potentially many), so don't let each call
+    // free+reload its bucket arenas (release: false). Instead free the model's
+    // bucket methods ONCE in the finally below, after the whole page — keeping
+    // the page's working set cached while still bounding memory across pages.
+    try {
+      // Mode A — no layout: OCR the whole page into one block.
+      if (!layout) {
+        const detections = ocr.runOCRWorklet(img, { release: false }).detections;
+        const blocks = detections.length
+          ? [
+              makeBlock<L>(
+                'ungrouped',
+                boundingBoxOf(detections.flatMap((d) => d.quad as Point[])),
+                1,
+                detections,
+                false
+              ),
+            ]
+          : [];
+        return { blocks, regions: [], detections, image: img };
       }
-      detections.push(...lines);
-      let block = makeBlock<L>(region.label, region.box, region.confidence, lines, table);
-      if (table && supporting) {
-        const structure = supporting.recognizeTableWorklet(crop);
-        block = { ...block, tableHtml: fillTableCells(structure.html, block.lines) };
+
+      // Mode B — layout: OCR each text region's crop on its own (upscaled into the
+      // detector → far better recall than one whole-page pass), offsetting lines
+      // back to page coords. Tables also recognize structure + fill cells.
+      const regions = layout.detectObjectsWorklet(img);
+      const blocks: DocumentBlock<L>[] = [];
+      const detections: OCRDetection[] = [];
+      for (const region of regions) {
+        if (!isTextRegion(region.label)) {
+          continue;
+        }
+        const { xmin, ymin } = region.box;
+        const crop = cropImageBuffer(img, region.box);
+        const lines = ocr
+          .runOCRWorklet(crop, { release: false })
+          .detections.map((d) => offsetDetection(d, xmin, ymin));
+        const table = isTableLabel(region.label);
+        if (lines.length === 0 && !table) {
+          continue;
+        }
+        detections.push(...lines);
+        let block = makeBlock<L>(region.label, region.box, region.confidence, lines, table);
+        if (table && supporting) {
+          const structure = supporting.recognizeTableWorklet(crop);
+          block = { ...block, tableHtml: fillTableCells(structure.html, block.lines) };
+        }
+        blocks.push(block);
       }
-      blocks.push(block);
+      blocks.sort((a, b) => a.bbox.ymin - b.bbox.ymin || a.bbox.xmin - b.bbox.xmin);
+      return { blocks, regions, detections, image: img };
+    } finally {
+      // Free the OCR model's bucket arenas once, after the whole page (the
+      // per-region runs used release: false). Bounds memory across pages while
+      // keeping each page's working set cached during the run.
+      ocr.releaseMethodsWorklet();
     }
-    blocks.sort((a, b) => a.bbox.ymin - b.bbox.ymin || a.bbox.xmin - b.bbox.xmin);
-    return { blocks, regions, detections, image: img };
   };
 
   const runDocumentOCR = wrapAsync(runDocumentOCRWorklet, runtime);
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 70a8f3ad6c..a9aed9a63e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -84,6 +84,15 @@ export type RunOCROptions = {
   readonly tallCropRatio?: number;
   /** Max stacked-column re-detection passes per page (each is detector-scale). Default 8. */
   readonly maxRedetections?: number;
+  /**
+   * Free the model's bucket-method activation arenas (`detect_<S>`/`recognize_<W>`)
+   * after this run, so memory doesn't accumulate as image/box sizes vary across
+   * runs (worse on CoreML, which compiles a graph per method). Default `true`.
+   * The document orchestrator passes `false` for its per-region OCR calls and
+   * frees once per page via `releaseMethods` instead, so it keeps the run's
+   * working set cached while still bounding memory.
+   */
+  readonly release?: boolean;
 };
 
 // The unified baked contract leaves only two things per detector architecture:
@@ -304,6 +313,31 @@ type VerticalContext = {
   readonly redetectBudget: { remaining: number };
 };
 
+// Divides an ordered TL,TR,BR,BL box into `parts` equal vertical bands (each a
+// TL,TR,BR,BL quad), top -> bottom. Used to recover the individual upright
+// letters of a stacked column from a box the detector merged (DBNet emits one
+// box per text region, not per glyph, so stacked letters arrive fused). `parts`
+// <= 1 returns the box unchanged.
+function splitTallQuad(ordered: readonly Point[], parts: number): Point[][] {
+  'worklet';
+  if (parts <= 1) {
+    return [ordered as Point[]];
+  }
+  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
+  const lerp = (a: Point, b: Point, t: number): Point => ({
+    x: a.x + (b.x - a.x) * t,
+    y: a.y + (b.y - a.y) * t,
+  });
+  const out: Point[][] = [];
+  for (let i = 0; i < parts; i++) {
+    const t0 = i / parts;
+    const t1 = (i + 1) / parts;
+    // Left edge runs tl->bl, right edge tr->br; take the band between t0 and t1.
+    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
+  }
+  return out;
+}
+
 // Joins glyph quads (in `src` pixel space, in reading order) into one recognizer
 // strip — each glyph warped upright to the recognizer height and laid side by
 // side — and recognizes it as a single line (joint hconcat). Returns null when
@@ -320,24 +354,34 @@ function recognizeGlyphStrip(
   'worklet';
   const recH = recCtx.recH;
   const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-  // Warp each glyph upright to recognizer height (aspect preserved).
+  // Warp each glyph upright to recognizer height (aspect preserved). A box that
+  // is much taller than wide is a merged run of stacked letters — split it into
+  // ~square single-letter cells first, so each lands in its own strip slot
+  // (otherwise N letters get squashed into one cell and read as garbage).
   const slices: { tGlyph: Tensor; w: number }[] = [];
   let totalW = 0;
   for (const g of glyphs) {
-    const gs = quadSize(g);
-    if (gs.width < 1 || gs.height < 1) {
+    const gsz = quadSize(g);
+    if (gsz.width < 1 || gsz.height < 1) {
       continue;
     }
-    const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
-    const tGlyph = tensor('uint8', [recH, gw, recC]);
-    warpQuad(src, tGlyph, flattenQuad(g), {
-      contentWidth: gw,
-      align: 'left',
-      padMode: 'constant',
-      padValue: RECOGNIZER_PAD_VALUE,
-    });
-    slices.push({ tGlyph, w: gw });
-    totalW += gw;
+    const parts = Math.max(1, Math.round(gsz.height / Math.max(1, gsz.width)));
+    for (const cell of splitTallQuad(g, parts)) {
+      const gs = quadSize(cell);
+      if (gs.width < 1 || gs.height < 1) {
+        continue;
+      }
+      const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
+      const tGlyph = tensor('uint8', [recH, gw, recC]);
+      warpQuad(src, tGlyph, flattenQuad(cell), {
+        contentWidth: gw,
+        align: 'left',
+        padMode: 'constant',
+        padValue: RECOGNIZER_PAD_VALUE,
+      });
+      slices.push({ tGlyph, w: gw });
+      totalW += gw;
+    }
   }
   if (slices.length === 0) {
     return null;
@@ -600,6 +644,10 @@ export async function createOCR(
   dispose: () => void;
   runOCR: (input: ImageBuffer, options?: RunOCROptions) => Promise<OCRResult>;
   runOCRWorklet: (input: ImageBuffer, options?: RunOCROptions) => OCRResult;
+  /** Free all bucket-method arenas without disposing the model (see `RunOCROptions.release`). */
+  releaseMethods: () => Promise<void>;
+  /** Worklet-thread variant of {@link releaseMethods}. */
+  releaseMethodsWorklet: () => void;
 }> {
   const { modelPath, ocrOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
@@ -706,11 +754,26 @@ export async function createOCR(
     model.dispose();
   };
 
+  // Free every per-size method's activation arena (detect_<S>/recognize_<W>)
+  // without disposing the model — they transparently reload on next execute.
+  // Defined before runOCRWorklet so the worklet plugin captures it (referenced
+  // worklets must precede their callers in source order).
+  const releaseMethodsWorklet = () => {
+    'worklet';
+    for (const s of detBuckets) {
+      model.unloadMethod(`detect_${s}`);
+    }
+    for (const w of recBuckets) {
+      model.unloadMethod(`recognize_${w}`);
+    }
+  };
+
   const runOCRWorklet = (input: ImageBuffer, options?: RunOCROptions): OCRResult => {
     'worklet';
     const vertical = options?.vertical ?? false;
     const tallCropRatio = options?.tallCropRatio ?? TALL_CROP_RATIO;
     const maxRedetections = options?.maxRedetections ?? MAX_VERTICAL_REDETECTIONS;
+    const release = options?.release ?? true;
     const { data, width, height, format } = input;
     const numChannels = FORMAT_CHANNELS[format];
     // Both detector and recognizer read RGB, so one conversion code serves both.
@@ -793,7 +856,10 @@ export async function createOCR(
       const { columns, singles } = groupVerticalColumns(ordered);
       for (const col of columns) {
         const boxStart = nowMs();
-        const r = recognizeGlyphStrip(recCtx, recSrc, recC, col); // col is top -> bottom
+        // `recognizeGlyphStrip` splits any multi-letter box into single-glyph
+        // cells (DBNet merges stacked letters and won't split them), so the
+        // column's boxes can be passed straight through, top -> bottom.
+        const r = recognizeGlyphStrip(recCtx, recSrc, recC, col);
         if (r) {
           pushVertical(r.text, r.conf, boundingQuadOf(col), nowMs() - boxStart);
         }
@@ -807,10 +873,17 @@ export async function createOCR(
     } finally {
       tInputRaw.dispose();
       tRecImage.dispose();
+      // Standalone runs free their bucket arenas so memory stays bounded as
+      // sizes vary; the document orchestrator opts out (release: false) and
+      // frees once per page.
+      if (release) {
+        releaseMethodsWorklet();
+      }
     }
   };
 
   const runOCR = wrapAsync(runOCRWorklet, runtime);
+  const releaseMethods = wrapAsync(releaseMethodsWorklet, runtime);
 
-  return { runOCR, runOCRWorklet, dispose };
+  return { runOCR, runOCRWorklet, dispose, releaseMethods, releaseMethodsWorklet };
 }

From 0621e5b64c886928d7852615508fa4f117af79a7 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 10:59:04 +0200
Subject: [PATCH 04/29] [RNE Rewrite] feat(ocr): reading-order sort +
 degenerate-dewarp guard

Reading order: add readingOrderIndices (column detection via x-coverage
sweep, within-column line grouping by vertical overlap, left-to-right
within a line, columns left-to-right). Apply to OCR detections and to each
document block's lines, replacing the detector's arbitrary / y-only order
so two-column pages, split titles, and label/value rows concatenate
correctly.

Dewarp guard: dewarpWorklet declines a degenerate warp (one that lacks
page boundaries and maps content off-canvas) by comparing sampled pixel
activity before/after; if the dewarped page keeps <50% of the source's
activity it returns the original, so dewarp can no longer collapse a page
to zero detections.
---
 .../src/extensions/cv/tasks/documentOCR.ts    |  18 +--
 .../src/extensions/cv/tasks/ocr.ts            |  18 ++-
 .../src/extensions/cv/tasks/ocrHelpers.ts     | 127 ++++++++++++++++++
 .../src/extensions/cv/tasks/supporting.ts     |  38 ++++++
 4 files changed, 187 insertions(+), 14 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
index ac2e5c2daa..dbc20661b6 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -11,6 +11,7 @@ import {
   type ObjectDetection,
 } from './objectDetection';
 import { createSupporting, type SupportingModel } from './supporting';
+import { readingOrderIndices } from './ocrHelpers';
 import { cropImageBuffer, rotateImageBuffer, fillTableCells } from './documentHelpers';
 
 /**
@@ -89,12 +90,6 @@ export type RunDocumentOCROptions = {
   readonly dewarp?: boolean;
 };
 
-const centerOf = (quad: readonly Point[]): { x: number; y: number } => {
-  'worklet';
-  const b = boundingBoxOf(quad);
-  return { x: (b.xmin + b.xmax) / 2, y: (b.ymin + b.ymax) / 2 };
-};
-
 const isTableLabel = (label: unknown): boolean => {
   'worklet';
   return String(label) === 'table';
@@ -121,12 +116,11 @@ function makeBlock<L>(
   isTable: boolean
 ): DocumentBlock<L> {
   'worklet';
-  // Decorate with each line's y-center once, then sort top-to-bottom (avoids
-  // recomputing the bbox inside the O(n log n) comparator).
-  const sorted = lines
-    .map((line) => ({ line, cy: centerOf(line.quad).y }))
-    .sort((a, b) => a.cy - b.cy)
-    .map((e) => e.line);
+  // Order the block's lines in reading order (top-to-bottom, and left-to-right
+  // within a line) so multi-column regions, titles split into words, and
+  // label/value rows concatenate correctly — not in the detector's arbitrary order.
+  const order = readingOrderIndices(lines.map((l) => l.quad));
+  const sorted = order.map((i) => lines[i]!);
   return {
     regionType,
     bbox,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index a9aed9a63e..78ea7a9b34 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -33,6 +33,7 @@ import {
   nowMs,
   snapDetectBucket,
   snapRecognizeBucket,
+  readingOrderIndices,
   type Buckets,
 } from './ocrHelpers';
 
@@ -493,6 +494,19 @@ function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
   ];
 }
 
+// Reorders recognized detections into human reading order (the detector emits
+// boxes in an arbitrary order). Column-aware: genuine multi-column pages read
+// column-by-column, single-column pages line-by-line, words within a line
+// left-to-right. Defined before its caller so the worklet plugin captures it.
+function orderDetections(dets: OCRDetection[]): OCRDetection[] {
+  'worklet';
+  if (dets.length <= 1) {
+    return dets;
+  }
+  const order = readingOrderIndices(dets.map((d) => d.quad));
+  return order.map((i) => dets[i]!);
+}
+
 // Clusters glyph-like, x-aligned, stacked boxes into vertical columns; wide lines
 // and isolated boxes come back as `singles` to read normally. So `vertical` ADDS
 // column reading without disturbing horizontal reads.
@@ -848,7 +862,7 @@ export async function createOCR(
           const { text, conf } = recognizeQuad(recCtx, recSrc, o);
           push(text, conf, o, nowMs() - boxStart);
         }
-        return { detections };
+        return { detections: orderDetections(detections) };
       }
 
       // Additive vertical pass: read x-aligned stacked glyph boxes as one joined
@@ -869,7 +883,7 @@ export async function createOCR(
         const { text, conf, stacked } = readBoxVertical(recCtx, vctx, recSrc, o, quadSize(o));
         (stacked ? pushVertical : push)(text, conf, o, nowMs() - boxStart);
       }
-      return { detections };
+      return { detections: orderDetections(detections) };
     } finally {
       tInputRaw.dispose();
       tRecImage.dispose();
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
index d8884b00c0..1faae5f8f2 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
@@ -157,6 +157,133 @@ export function quadSize(ordered: readonly Point[]): { width: number; height: nu
   return { width, height };
 }
 
+// Reading-order tuning. A vertical gutter crossed by no box must be at least this
+// fraction of the content width to separate columns; two boxes share a text line
+// when their vertical extents overlap by at least this fraction of the shorter
+// box's height. Stable across models — not worth exposing.
+const READING_COL_GAP_FRAC = 0.06;
+const READING_ROW_OVERLAP_FRAC = 0.3;
+
+type ReadingBox = { xmin: number; ymin: number; xmax: number; ymax: number };
+
+// Axis-aligned bounds of a quad. Defined before readingOrderIndices so the worklet
+// plugin captures it (a referenced worklet must precede its caller in source order).
+function bboxOfQuad(quad: readonly Point[]): ReadingBox {
+  'worklet';
+  let xmin = Infinity;
+  let ymin = Infinity;
+  let xmax = -Infinity;
+  let ymax = -Infinity;
+  for (const p of quad) {
+    if (p.x < xmin) xmin = p.x;
+    if (p.y < ymin) ymin = p.y;
+    if (p.x > xmax) xmax = p.x;
+    if (p.y > ymax) ymax = p.y;
+  }
+  return { xmin, ymin, xmax, ymax };
+}
+
+/**
+ * Returns the indices of `quads` in human reading order. Detects vertical column
+ * gutters with an x-coverage sweep — a band crossed by no box and wider than
+ * {@link READING_COL_GAP_FRAC} of the content width splits columns (within-line
+ * word gaps don't, because other lines cover that x); within each column it groups
+ * boxes into lines by vertical overlap, orders lines top-to-bottom and the boxes in
+ * a line left-to-right, and reads columns left-to-right. Single-line / single-column
+ * inputs collapse to the obvious order. The detector emits boxes in an arbitrary
+ * order, so callers reorder their detections/lines through this.
+ * @category Typescript API
+ * @param quads The detected text quads (any corner order).
+ * @returns Indices into `quads`, in reading order.
+ */
+export function readingOrderIndices(quads: readonly (readonly Point[])[]): number[] {
+  'worklet';
+  const n = quads.length;
+  if (n <= 1) {
+    return n === 1 ? [0] : [];
+  }
+  const boxes = quads.map((q) => bboxOfQuad(q));
+  let X0 = Infinity;
+  let X1 = -Infinity;
+  for (const b of boxes) {
+    if (b.xmin < X0) X0 = b.xmin;
+    if (b.xmax > X1) X1 = b.xmax;
+  }
+  const minGap = READING_COL_GAP_FRAC * Math.max(1, X1 - X0);
+
+  // Column detection: sweep box x-edges; an internal span where coverage drops to
+  // zero (no box at that x across any line) wider than minGap is a column gutter.
+  const edges: { x: number; d: number }[] = [];
+  for (const b of boxes) {
+    edges.push({ x: b.xmin, d: 1 });
+    edges.push({ x: b.xmax, d: -1 });
+  }
+  // At equal x, open (+1) before close (-1) so touching boxes don't open a gutter.
+  edges.sort((a, b) => a.x - b.x || b.d - a.d);
+  const cuts: number[] = [];
+  let active = 0;
+  let gutterStart = 0;
+  for (const e of edges) {
+    const before = active;
+    active += e.d;
+    if (before > 0 && active === 0) {
+      gutterStart = e.x;
+    } else if (before === 0 && active > 0 && e.x - gutterStart >= minGap) {
+      cuts.push((gutterStart + e.x) / 2);
+    }
+  }
+
+  // Bucket boxes into columns by center-x relative to the (ascending) cut lines.
+  const numCols = cuts.length + 1;
+  const columns: number[][] = [];
+  for (let c = 0; c < numCols; c++) {
+    columns.push([]);
+  }
+  for (let i = 0; i < n; i++) {
+    const cx = (boxes[i]!.xmin + boxes[i]!.xmax) / 2;
+    let c = 0;
+    for (const cut of cuts) {
+      if (cx > cut) c++;
+    }
+    columns[c]!.push(i);
+  }
+
+  // Within each column: group boxes into lines by vertical overlap, order lines
+  // top-to-bottom and boxes within a line left-to-right.
+  const out: number[] = [];
+  for (const col of columns) {
+    col.sort((a, b) => boxes[a]!.ymin - boxes[b]!.ymin);
+    const rows: { items: number[]; ymin: number; ymax: number }[] = [];
+    for (const i of col) {
+      const b = boxes[i]!;
+      let placed = false;
+      for (const row of rows) {
+        const overlap = Math.min(row.ymax, b.ymax) - Math.max(row.ymin, b.ymin);
+        const minH = Math.min(row.ymax - row.ymin, b.ymax - b.ymin);
+        if (overlap >= READING_ROW_OVERLAP_FRAC * Math.max(1, minH)) {
+          row.items.push(i);
+          row.ymin = Math.min(row.ymin, b.ymin);
+          row.ymax = Math.max(row.ymax, b.ymax);
+          placed = true;
+          break;
+        }
+      }
+      if (!placed) {
+        rows.push({ items: [i], ymin: b.ymin, ymax: b.ymax });
+      }
+    }
+    rows.sort((a, b) => a.ymin - b.ymin);
+    const cx = (i: number): number => boxes[i]!.xmin + boxes[i]!.xmax;
+    for (const row of rows) {
+      row.items.sort((a, b) => cx(a) - cx(b));
+      for (const i of row.items) {
+        out.push(i);
+      }
+    }
+  }
+  return out;
+}
+
 /**
  * Resolves the content width (px) of a recognizer crop: the quad resized to the
  * recognizer height keeping aspect, clamped to the model's bucket width.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
index 5a54d01d71..d39b08918b 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
@@ -16,6 +16,37 @@ import { argmaxRange } from './documentHelpers';
 const DEFAULT_EOS_TOKEN_ID = 49;
 const DEFAULT_MAX_STEPS = 501; // SLANet max_text_length (500) + 1
 
+// A dewarp grid estimated on a page without clear boundaries (e.g. text floating
+// on white, or a clean scan UVDoc wasn't trained on) can map most of the output
+// off the source, collapsing the page to near-blank and OCR to zero detections.
+// Guard: if the dewarped page keeps less than this fraction of the source's pixel
+// activity (variance), the warp is degenerate — keep the original page.
+const DEWARP_MIN_ACTIVITY_RATIO = 0.5;
+const DEWARP_ACTIVITY_STRIDE = 31; // subsample every Nth pixel for a cheap estimate
+
+// Variance of one channel sampled every DEWARP_ACTIVITY_STRIDE pixels — a cheap,
+// polarity-independent proxy for how much content (ink/edges) an image carries. A
+// blank/uniform page is ~0. Defined before createSupporting so the worklet plugin
+// captures it (a referenced worklet must precede its caller in source order).
+function sampledActivity(data: Uint8Array, channels: number): number {
+  'worklet';
+  let n = 0;
+  let sum = 0;
+  let sumSq = 0;
+  const step = channels * DEWARP_ACTIVITY_STRIDE;
+  for (let i = 0; i < data.length; i += step) {
+    const v = data[i]!;
+    sum += v;
+    sumSq += v * v;
+    n++;
+  }
+  if (n === 0) {
+    return 0;
+  }
+  const mean = sum / n;
+  return sumSq / n - mean * mean;
+}
+
 /**
  * Detected page orientation.
  * @category Types
@@ -213,6 +244,13 @@ export async function createSupporting(
       gridSample(tSrc, tGrid, tDst);
       const out = new Uint8Array(input.width * input.height * ch);
       tDst.getData(out);
+      // Degenerate-warp guard: a grid lacking page boundaries can push content
+      // off-canvas, leaving a near-blank page. If the dewarp collapsed the image's
+      // activity, decline it and keep the original (better an un-dewarped read than
+      // zero detections).
+      if (sampledActivity(out, ch) < DEWARP_MIN_ACTIVITY_RATIO * sampledActivity(input.data, ch)) {
+        return input;
+      }
       return {
         data: out,
         width: input.width,

From 9a92ce336b75e048da258f017e1fa0e8a6dafd0d Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 11:10:10 +0200
Subject: [PATCH 05/29] [RNE Rewrite] debug(ocr): log orientation head output

Warn the raw orientation logits (per-class), the argmax, the decoded
rotationCW and confidence from detectOrientationWorklet. console.warn so
it surfaces in native logs from the worklet thread.
---
 .../src/extensions/cv/tasks/supporting.ts           | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
index d39b08918b..5a17fd8a85 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
@@ -227,7 +227,18 @@ export async function createSupporting(
     for (let i = 0; i < oriOutLen; i++) {
       sumExp += Math.exp(oriBuf[i]! - best);
     }
-    return { rotationCW: ((cls % 4) * 90) as 0 | 90 | 180 | 270, confidence: 1 / sumExp };
+    const rotationCW = ((cls % 4) * 90) as 0 | 90 | 180 | 270;
+    const confidence = 1 / sumExp;
+    // DEBUG: raw orientation head output (logits per class) + decoded result.
+    // console.warn (not console.log) so it surfaces in native logs from the worklet.
+    let logitsStr = '';
+    for (let i = 0; i < oriOutLen; i++) {
+      logitsStr += (i === 0 ? '' : ', ') + oriBuf[i]!.toFixed(3);
+    }
+    console.warn(
+      `[orientation] logits=[${logitsStr}] argmax=${cls} rotationCW=${rotationCW} confidence=${confidence.toFixed(3)}`
+    );
+    return { rotationCW, confidence };
   };
 
   const dewarpWorklet = (input: ImageBuffer): ImageBuffer => {

From 02f748de0a110cb53904df6592fab4171740b30f Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 12:40:06 +0200
Subject: [PATCH 06/29] [RNE Rewrite] fix(ocr): gate orientation correction on
 confidence

Only apply page-rotation when the orientation classifier's softmax
confidence for its argmax class is >= 0.7 and the predicted angle is
non-zero, mirroring PaddleOCR's pipeline. Out-of-distribution inputs
(perspective photos, non-documents) produce low-confidence argmaxes that
spuriously flip the page; below threshold the page is treated as upright.
---
 .../src/extensions/cv/tasks/documentOCR.ts        | 15 ++++++++++++---
 1 file changed, 12 insertions(+), 3 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
index dbc20661b6..7eff47323e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -90,6 +90,12 @@ export type RunDocumentOCROptions = {
   readonly dewarp?: boolean;
 };
 
+// Minimum orientation-classifier confidence (softmax of the argmax class) to act
+// on a non-zero rotation. Mirrors PaddleOCR's pipeline gate: out-of-distribution
+// inputs (photos, non-documents) produce low-confidence argmaxes that spuriously
+// flip the page, so below this we treat the page as already upright (0°).
+const ORIENTATION_MIN_CONFIDENCE = 0.7;
+
 const isTableLabel = (label: unknown): boolean => {
   'worklet';
   return String(label) === 'table';
@@ -179,9 +185,12 @@ export async function createDocumentOCR<L>(
     const useDewarp = !!supporting && (options?.dewarp ?? defaultDewarp);
     let img = input;
     if (useOrientation && supporting) {
-      const rot = supporting.detectOrientationWorklet(img).rotationCW;
-      if (rot !== 0) {
-        img = rotateImageBuffer(img, ((360 - rot) % 360) as 0 | 90 | 180 | 270);
+      // Only correct when the classifier is confident AND the predicted angle is
+      // non-zero — a low-confidence argmax (typical of OOD photos / non-documents)
+      // otherwise spuriously flips the page. Below threshold we leave it as 0°.
+      const ori = supporting.detectOrientationWorklet(img);
+      if (ori.rotationCW !== 0 && ori.confidence >= ORIENTATION_MIN_CONFIDENCE) {
+        img = rotateImageBuffer(img, ((360 - ori.rotationCW) % 360) as 0 | 90 | 180 | 270);
       }
     }
     if (useDewarp && supporting) {

From f6f30e664b136aac75616d4ed0686fcdced17ed1 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 12:41:34 +0200
Subject: [PATCH 07/29] [RNE Rewrite] fix(ocr): raise orientation confidence
 gate to 0.85

Genuine documents score >0.95; OOD frames can land ~0.74, so 0.85 leaves
margin to reject the spurious flips a 0.7 gate let through.
---
 .../src/extensions/cv/tasks/documentOCR.ts                  | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
index 7eff47323e..7de026f093 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -93,8 +93,10 @@ export type RunDocumentOCROptions = {
 // Minimum orientation-classifier confidence (softmax of the argmax class) to act
 // on a non-zero rotation. Mirrors PaddleOCR's pipeline gate: out-of-distribution
 // inputs (photos, non-documents) produce low-confidence argmaxes that spuriously
-// flip the page, so below this we treat the page as already upright (0°).
-const ORIENTATION_MIN_CONFIDENCE = 0.7;
+// flip the page, so below this we treat the page as already upright (0°). Set high
+// (0.85) — genuine documents score >0.95, leaving margin to reject OOD frames that
+// can still land ~0.74.
+const ORIENTATION_MIN_CONFIDENCE = 0.85;
 
 const isTableLabel = (label: unknown): boolean => {
   'worklet';

From 20fb68365115c4420a48f70abbb9be039d28743a Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 14:40:14 +0200
Subject: [PATCH 08/29] [RNE Rewrite] refactor(ocr): tidy useDocumentOCR
 local-config assembly

Factor a localize() helper that swaps a model spec's hosted modelPath for
its downloaded local path (undefined when the optional model is absent or
not yet downloaded). Replaces the nested conditional-spread localConfig
with a flat object, and aggregates progress/error over just the enabled
downloads. Behavior unchanged.
---
 .../src/hooks/useDocumentOCR.ts               | 43 +++++++++++--------
 1 file changed, 24 insertions(+), 19 deletions(-)

diff --git a/packages/react-native-executorch/src/hooks/useDocumentOCR.ts b/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
index a48c11d6a4..a1e3bd87f5 100644
--- a/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
+++ b/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
@@ -2,6 +2,14 @@ import { useModel } from './useModel';
 import { useResourceDownload } from './useResourceDownload';
 import { createDocumentOCR, type DocumentOCRModel } from '../extensions/cv/tasks/documentOCR';
 
+// Swap a model spec's hosted `modelPath` for its downloaded local path. Returns
+// undefined when the spec is absent (an optional model) or its path hasn't
+// finished downloading yet.
+const localize = <M extends { modelPath: string }>(
+  spec: M | undefined,
+  localPath: string | undefined
+): M | undefined => (spec && localPath ? { ...spec, modelPath: localPath } : undefined);
+
 /**
  * React hook for the document OCR pipeline: OCR + optional layout detection +
  * optional supporting (orientation/dewarp/table), assembled into reading-ordered
@@ -22,19 +30,14 @@ export function useDocumentOCR<L>(
   const layoutDl = useResourceDownload(config.layout?.modelPath, options?.preventLoad);
   const supDl = useResourceDownload(config.supporting?.modelPath, options?.preventLoad);
 
-  const ready =
-    !!ocrDl.localPath &&
-    (!config.layout || !!layoutDl.localPath) &&
-    (!config.supporting || !!supDl.localPath);
+  // Localize each enabled model; an optional model is "ready" when it's either
+  // absent or fully downloaded. Build the run config only once all are ready.
+  const ocr = localize(config.ocr, ocrDl.localPath);
+  const layout = localize(config.layout, layoutDl.localPath);
+  const supporting = localize(config.supporting, supDl.localPath);
+  const ready = !!ocr && (!config.layout || !!layout) && (!config.supporting || !!supporting);
   const localConfig: DocumentOCRModel<L> | null = ready
-    ? {
-        ...config,
-        ocr: { ...config.ocr, modelPath: ocrDl.localPath! },
-        ...(config.layout ? { layout: { ...config.layout, modelPath: layoutDl.localPath! } } : {}),
-        ...(config.supporting
-          ? { supporting: { ...config.supporting, modelPath: supDl.localPath! } }
-          : {}),
-      }
+    ? { ...config, ocr: ocr!, layout, supporting }
     : null;
 
   const { model, error } = useModel(createDocumentOCR<L>, localConfig, [
@@ -43,16 +46,18 @@ export function useDocumentOCR<L>(
     supDl.localPath,
   ]);
 
-  // Overall progress is the slowest of the enabled downloads, so it can't read
-  // 100% while a second/third model is still fetching.
-  const progresses = [ocrDl.downloadProgress];
-  if (config.layout) progresses.push(layoutDl.downloadProgress);
-  if (config.supporting) progresses.push(supDl.downloadProgress);
+  // Aggregate only the ENABLED downloads, so progress can't read 100% while a
+  // second/third model is still fetching.
+  const downloads = [
+    ocrDl,
+    ...(config.layout ? [layoutDl] : []),
+    ...(config.supporting ? [supDl] : []),
+  ];
 
   return {
     isReady: !!model,
-    error: ocrDl.downloadError || layoutDl.downloadError || supDl.downloadError || error,
-    downloadProgress: Math.min(...progresses),
+    error: downloads.map((d) => d.downloadError).find(Boolean) || error,
+    downloadProgress: Math.min(...downloads.map((d) => d.downloadProgress)),
     runDocumentOCR: model?.runDocumentOCR,
     runDocumentOCRWorklet: model?.runDocumentOCRWorklet,
   };

From 82ffdbceb6b0482f1be7b4221553ed29bf5a5138 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:10:34 +0200
Subject: [PATCH 09/29] [RNE Rewrite] feat(ocr): configurable recognizer norm +
 pluggable decode

OCROptions gains recognizerNorm (alpha/beta), recognizerPadValue, and an
optional decode(logits, charset) -> {text, confidence}. Defaults preserve
the SVTR/CRNN contract ((x/255-0.5)/0.5, pad 128, greedy CTC), so existing
models are unchanged; a model with different normalization or a non-CTC
head (attention/AR) now slots in as pure config. decode runs on the
worklet thread (must be a worklet).
---
 .../src/extensions/cv/tasks/ocr.ts            | 63 +++++++++++++++++--
 1 file changed, 57 insertions(+), 6 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 78ea7a9b34..5bba4b33da 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -65,6 +65,29 @@ export type OCROptions = {
   readonly charset: string | readonly string[];
   /** Drop detections below this confidence. Defaults per detector architecture. */
   readonly dropScore?: number;
+  /**
+   * Recognizer input normalization, applied after the warp as `x·alpha + beta`
+   * (scalar, or per-RGB-channel `[r,g,b]`). Defaults to `(x/255 − 0.5)/0.5` →
+   * `[−1,1]` (`alpha = 1/127.5`, `beta = −1`), the SVTR/CRNN convention. Override
+   * for a recognizer trained with different normalization (e.g. ImageNet).
+   */
+  readonly recognizerNorm?: {
+    readonly alpha: number | readonly number[];
+    readonly beta: number | readonly number[];
+  };
+  /** Fill value for the recognizer canvas padding. Defaults to 128 (neutral gray). */
+  readonly recognizerPadValue?: number;
+  /**
+   * Custom recognizer decode, replacing the built-in greedy CTC. Receives the
+   * raw `recognize_<W>` output tensor (shape `[1, T, V]`, softmaxed per the
+   * contract) and the charset, and returns the recognized text plus a confidence
+   * in `[0,1]`. Use for non-CTC heads (attention/AR decoders) or custom scoring.
+   * MUST be a worklet — it runs on the pipeline's worklet thread.
+   */
+  readonly decode?: (
+    logits: Tensor,
+    charset: readonly string[]
+  ) => { readonly text: string; readonly confidence: number };
 };
 
 /**
@@ -267,6 +290,15 @@ type RecContext = {
   readonly recBuckets: readonly number[];
   readonly recH: number;
   readonly charset: string[];
+  // Per-model recognizer normalization / pad (resolved from OCROptions defaults).
+  readonly normAlpha: number | readonly number[];
+  readonly normBeta: number | readonly number[];
+  readonly padValue: number;
+  // Optional custom decode; falls back to greedy CTC when absent.
+  readonly decode?: (
+    logits: Tensor,
+    charset: readonly string[]
+  ) => { readonly text: string; readonly confidence: number };
 };
 
 // Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap content width to a
@@ -288,14 +320,19 @@ function recognizeQuad(
     contentWidth,
     align: 'left',
     padMode: 'constant',
-    padValue: RECOGNIZER_PAD_VALUE,
+    padValue: ctx.padValue,
   });
   rs.tCanvas
     .through(toChannelsFirst, rs.tCF)
-    .through(normalize, rs.tNorm, { alpha: RECOGNIZER_ALPHA, beta: RECOGNIZER_BETA })
+    .through(normalize, rs.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
     .copyTo(rs.tInput);
   ctx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
-  // Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
+  // A custom decode (e.g. attention/AR head) takes the raw logits; otherwise
+  // greedy CTC. Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
+  if (ctx.decode) {
+    const r = ctx.decode(rs.tLogits, ctx.charset);
+    return { text: r.text, conf: r.confidence };
+  }
   const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
   const text = decodeGreedy(indices, ctx.charset);
   const conf = ctcConfidence(values, indices);
@@ -378,7 +415,7 @@ function recognizeGlyphStrip(
         contentWidth: gw,
         align: 'left',
         padMode: 'constant',
-        padValue: RECOGNIZER_PAD_VALUE,
+        padValue: recCtx.padValue,
       });
       slices.push({ tGlyph, w: gw });
       totalW += gw;
@@ -396,7 +433,7 @@ function recognizeGlyphStrip(
     const rs = recCtx.recSetByWidth.get(bucketW)!;
     // Assemble the strip row-major into the bucket canvas, neutral-padded.
     const strip = new Uint8Array(recH * bucketW * recC);
-    strip.fill(RECOGNIZER_PAD_VALUE);
+    strip.fill(recCtx.padValue);
     let xOff = 0;
     for (const s of slices) {
       if (xOff >= bucketW) {
@@ -415,9 +452,13 @@ function recognizeGlyphStrip(
     rs.tCanvas.setData(strip);
     rs.tCanvas
       .through(toChannelsFirst, rs.tCF)
-      .through(normalize, rs.tNorm, { alpha: RECOGNIZER_ALPHA, beta: RECOGNIZER_BETA })
+      .through(normalize, rs.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
       .copyTo(rs.tInput);
     recCtx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+    if (recCtx.decode) {
+      const r = recCtx.decode(rs.tLogits, recCtx.charset);
+      return r.text.length > 0 ? { text: r.text, conf: r.confidence } : null;
+    }
     const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
     const text = decodeGreedy(indices, recCtx.charset);
     const conf = ctcConfidence(values, indices);
@@ -667,6 +708,12 @@ export async function createOCR(
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
   const dropScore = ocrOpts.dropScore ?? DEFAULT_DROP_SCORE[ocrOpts.detectorKind];
+  // Recognizer normalization / pad / decode — defaults preserve the SVTR/CRNN
+  // contract; OCROptions can override per model (see RecContext).
+  const recNormAlpha = ocrOpts.recognizerNorm?.alpha ?? RECOGNIZER_ALPHA;
+  const recNormBeta = ocrOpts.recognizerNorm?.beta ?? RECOGNIZER_BETA;
+  const recPadValue = ocrOpts.recognizerPadValue ?? RECOGNIZER_PAD_VALUE;
+  const recDecode = ocrOpts.decode;
 
   const detBuckets = ocrOpts.buckets.detect;
   const recBuckets = ocrOpts.buckets.recognize;
@@ -824,6 +871,10 @@ export async function createOCR(
         recBuckets,
         recH,
         charset,
+        normAlpha: recNormAlpha,
+        normBeta: recNormBeta,
+        padValue: recPadValue,
+        decode: recDecode,
       };
       // The vertical path crops each box from the raw page and re-detects its
       // characters; `recCode`/`recC` convert a box crop to RGB.

From 5fae2c1c9b3d6219ad3cfd753593557c485ada45 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:12:27 +0200
Subject: [PATCH 10/29] [RNE Rewrite] chore(ocr): remove orientation debug
 logging

Drop the per-call console.warn of orientation logits added for the OOD
investigation; the confidence gate is the shipping fix.
---
 .../src/extensions/cv/tasks/supporting.ts                | 9 ---------
 1 file changed, 9 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
index 5a17fd8a85..8c70702ba7 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
@@ -229,15 +229,6 @@ export async function createSupporting(
     }
     const rotationCW = ((cls % 4) * 90) as 0 | 90 | 180 | 270;
     const confidence = 1 / sumExp;
-    // DEBUG: raw orientation head output (logits per class) + decoded result.
-    // console.warn (not console.log) so it surfaces in native logs from the worklet.
-    let logitsStr = '';
-    for (let i = 0; i < oriOutLen; i++) {
-      logitsStr += (i === 0 ? '' : ', ') + oriBuf[i]!.toFixed(3);
-    }
-    console.warn(
-      `[orientation] logits=[${logitsStr}] argmax=${cls} rotationCW=${rotationCW} confidence=${confidence.toFixed(3)}`
-    );
     return { rotationCW, confidence };
   };
 

From 886ba883376788ace27a531666371915d066c215 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:22:32 +0200
Subject: [PATCH 11/29] [RNE Rewrite] refactor(cv): unify resize factors +
 guard degenerate inputs

- Extract resizeFactors() (points.ts) so scalePoint and scaleBox derive the
  letterbox/stretch scale+offset once instead of each recomputing it (#6).
- boundingBoxOf / bboxOfQuad / boundingQuadOf return a zero box/quad for
  empty input instead of Infinity bounds (#11).
- orderQuad returns a copy unchanged when not given exactly 4 corners (#12).
---
 .../src/extensions/cv/ops/boxes.ts            | 22 +++------
 .../src/extensions/cv/ops/points.ts           | 45 +++++++++++++------
 .../src/extensions/cv/tasks/ocr.ts            |  8 ++++
 .../src/extensions/cv/tasks/ocrHelpers.ts     |  8 ++++
 4 files changed, 52 insertions(+), 31 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index a783ce3443..361f69bc20 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,7 +1,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
-import { scalePoint, type Point } from './points';
+import { scalePoint, resizeFactors, type Point } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
@@ -73,22 +73,7 @@ export function scaleBox<F extends BoxFormat>(
   }
 ): BoundingBox<F> {
   'worklet';
-  const { from, to, resizeMode } = opts;
-
-  let scaleX: number;
-  let scaleY: number;
-  switch (resizeMode) {
-    case 'letterbox': {
-      const scale = Math.min(from.width / to.width, from.height / to.height);
-      scaleX = scale;
-      scaleY = scale;
-      break;
-    }
-    case 'stretch':
-      scaleX = from.width / to.width;
-      scaleY = from.height / to.height;
-      break;
-  }
+  const { scaleX, scaleY } = resizeFactors(opts.from, opts.to, opts.resizeMode);
 
   switch (box.format) {
     case 'xyxy': {
@@ -134,6 +119,9 @@ export function scaleBox<F extends BoxFormat>(
  */
 export function boundingBoxOf(points: readonly Point[]): BoundingBox<'xyxy'> {
   'worklet';
+  if (points.length === 0) {
+    return { format: 'xyxy', xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
+  }
   let xmin = Infinity;
   let ymin = Infinity;
   let xmax = -Infinity;
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 7114ffe04b..524c2b7f25 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -34,6 +34,35 @@ export function clamp(v: number, lo: number, hi: number): number {
  * 'stretch').
  * @returns The scaled coordinate point.
  */
+/**
+ * Per-axis scale and offset that map a `to`-space coordinate back into
+ * `from`-space, for an aspect-preserving letterbox or an axis stretch. The
+ * inverse map is `(coord − offset) / scale` per axis (offset is 0 for stretch).
+ * Shared by {@link scalePoint} and `scaleBox` so the factors are derived once.
+ * @category Utils
+ * @param from The source bounds (e.g. model input dimensions).
+ * @param to The destination bounds (e.g. original image dimensions).
+ * @param resizeMode The resize mode the source was produced with.
+ * @returns The per-axis `scaleX`/`scaleY` and `offsetX`/`offsetY`.
+ */
+export function resizeFactors(
+  from: { readonly width: number; readonly height: number },
+  to: { readonly width: number; readonly height: number },
+  resizeMode: Exclude<ResizeMode, 'crop'>
+): { scaleX: number; scaleY: number; offsetX: number; offsetY: number } {
+  'worklet';
+  if (resizeMode === 'letterbox') {
+    const scale = Math.min(from.width / to.width, from.height / to.height);
+    return {
+      scaleX: scale,
+      scaleY: scale,
+      offsetX: (from.width - to.width * scale) / 2.0,
+      offsetY: (from.height - to.height * scale) / 2.0,
+    };
+  }
+  return { scaleX: from.width / to.width, scaleY: from.height / to.height, offsetX: 0, offsetY: 0 };
+}
+
 export function scalePoint(
   point: Point,
   opts: {
@@ -43,18 +72,6 @@ export function scalePoint(
   }
 ): Point {
   'worklet';
-  const { from, to, resizeMode } = opts;
-  switch (resizeMode) {
-    case 'letterbox': {
-      const scale = Math.min(from.width / to.width, from.height / to.height);
-      const offsetX = (from.width - to.width * scale) / 2.0;
-      const offsetY = (from.height - to.height * scale) / 2.0;
-      return { x: (point.x - offsetX) / scale, y: (point.y - offsetY) / scale };
-    }
-    case 'stretch': {
-      const scaleX = from.width / to.width;
-      const scaleY = from.height / to.height;
-      return { x: point.x / scaleX, y: point.y / scaleY };
-    }
-  }
+  const { scaleX, scaleY, offsetX, offsetY } = resizeFactors(opts.from, opts.to, opts.resizeMode);
+  return { x: (point.x - offsetX) / scaleX, y: (point.y - offsetY) / scaleY };
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 5bba4b33da..3a73c8498f 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -515,6 +515,14 @@ function readStackedColumn(
 // Bounding axis-aligned quad (TL,TR,BR,BL) enclosing a set of quads.
 function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
   'worklet';
+  if (quads.length === 0) {
+    return [
+      { x: 0, y: 0 },
+      { x: 0, y: 0 },
+      { x: 0, y: 0 },
+      { x: 0, y: 0 },
+    ];
+  }
   let xmin = Infinity;
   let ymin = Infinity;
   let xmax = -Infinity;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
index 1faae5f8f2..9a5ed12b37 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
@@ -107,6 +107,11 @@ export function mapQuadToImage(
  */
 export function orderQuad(pts: readonly Point[]): Point[] {
   'worklet';
+  // The x±y-extreme ordering below assumes exactly four corners; for any other
+  // count return a copy unchanged rather than indexing out of range.
+  if (pts.length !== 4) {
+    return [...pts];
+  }
   let tl = 0;
   let tr = 0;
   let br = 0;
@@ -170,6 +175,9 @@ type ReadingBox = { xmin: number; ymin: number; xmax: number; ymax: number };
 // plugin captures it (a referenced worklet must precede its caller in source order).
 function bboxOfQuad(quad: readonly Point[]): ReadingBox {
   'worklet';
+  if (quad.length === 0) {
+    return { xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
+  }
   let xmin = Infinity;
   let ymin = Infinity;
   let xmax = -Infinity;

From 0803fdede9f328cdcb421b809c33a4a913aada4e Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:27:14 +0200
Subject: [PATCH 12/29] [RNE Rewrite] perf(ocr): cache detector scratch tensors
 per bucket
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

detectQuads allocated 7 tensors per call (~30MB at the 960 bucket), freed
on return — wasteful on the vertical re-detect path that calls it per box.
Pre-allocate the channel-independent set per detect bucket at construction
(buildDetectorSets, mirroring buildRecognizerSets); detectQuads now only
allocates the source-resize tensor (the lone input-channel-dependent one).
Behavior unchanged; disposed alongside recSets.
---
 .../src/extensions/cv/tasks/ocr.ts            | 94 +++++++++++++------
 1 file changed, 65 insertions(+), 29 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 3a73c8498f..6f820094fc 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -204,17 +204,30 @@ function toRgbCode(format: ImageFormat): ColorConversionCode | null {
 // Stored at construction; CRAFT's `targetHeight` is per-run, so detectQuads adds it.
 type DetectorExtractConfig = Omit<CraftExtractOptions, 'targetHeight'> | DbnetExtractOptions;
 
+// Per-detect-bucket scratch tensors, allocated once at construction (like RecSet)
+// and reused across the page pass and per-box re-detects. Only the source-resize
+// tensor depends on the run's input channel count, so detectQuads allocates that
+// one per call and reuses the rest from here.
+type DetSet = {
+  readonly s: number;
+  readonly tColor: Tensor; // [s, s, 3]
+  readonly tCF: Tensor; // [3, s, s]
+  readonly tNorm: Tensor; // [3, s, s]
+  readonly tInput: Tensor; // [1, 3, s, s]
+  readonly tHeatmap: Tensor; // dbnet [1,1,s,s] | craft [1, s/2, s/2, 2]
+  readonly tExtras: readonly Tensor[]; // craft extras at half-res
+};
+
 // Everything the detector pass needs, bundled so it can run both on the full
 // page and (for vertical text) on a single box crop to find its characters.
 type DetectContext = {
   readonly model: Model;
-  readonly detectorKind: 'craft' | 'dbnet';
   readonly detBuckets: readonly number[];
   readonly format: ImageFormat;
   readonly numChannels: number;
   readonly detCode: ColorConversionCode | null;
   readonly extractOpts: DetectorExtractConfig;
-  readonly detExtraChannels: readonly number[];
+  readonly detSets: ReadonlyMap<number, DetSet>;
 };
 
 // Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
@@ -229,45 +242,29 @@ function detectQuads(
 ): Point[][] {
   'worklet';
   const detS = snapDetectBucket(width, height, ctx.detBuckets);
-  // CRAFT heatmap is half-resolution; detector buckets are validated even.
-  const heat = detS / 2;
-
+  // snapDetectBucket always returns one of detBuckets, so the set exists.
+  const ds = ctx.detSets.get(detS)!;
+  // Only the source resize depends on the run's channel count; the rest is cached.
   const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
-  const tDetColor = tensor('uint8', [detS, detS, 3]);
-  const tDetCF = tensor('uint8', [3, detS, detS]);
-  const tDetNorm = tensor('float32', [3, detS, detS]);
-  const tDetInput = tensor('float32', [1, 3, detS, detS]);
-  // CRAFT: half-res [1,S/2,S/2,2] heatmap; DBNet: full-res [1,1,S,S] prob map.
-  const tHeatmap =
-    ctx.detectorKind === 'dbnet'
-      ? tensor('float32', [1, 1, detS, detS])
-      : tensor('float32', [1, heat, heat, 2]);
-  const tDetExtras = ctx.detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]));
   try {
     src
       .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
-      .throughIf(ctx.detCode !== null, cvtColor, tDetColor, ctx.detCode!)
-      .through(toChannelsFirst, tDetCF)
-      .through(normalize, tDetNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
-      .copyTo(tDetInput);
+      .throughIf(ctx.detCode !== null, cvtColor, ds.tColor, ctx.detCode!)
+      .through(toChannelsFirst, ds.tCF)
+      .through(normalize, ds.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
+      .copyTo(ds.tInput);
 
-    ctx.model.execute(`detect_${detS}`, [tDetInput], [tHeatmap, ...tDetExtras]);
+    ctx.model.execute(`detect_${detS}`, [ds.tInput], [ds.tHeatmap, ...ds.tExtras]);
     // CRAFT needs the per-run input height to restore its half-res boxes;
     // `charLevel` switches it to per-glyph (ungrouped) boxes for a column pass.
     const extractOpts =
       ctx.extractOpts.mode === 'craft'
         ? { ...ctx.extractOpts, targetHeight: detS, charLevel }
         : ctx.extractOpts;
-    const quads = extractTextBoxes(tHeatmap, extractOpts);
+    const quads = extractTextBoxes(ds.tHeatmap, extractOpts);
     return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
   } finally {
     tDetResize.dispose();
-    tDetColor.dispose();
-    tDetCF.dispose();
-    tDetNorm.dispose();
-    tDetInput.dispose();
-    tHeatmap.dispose();
-    tDetExtras.forEach((t) => t.dispose());
   }
 }
 
@@ -652,6 +649,31 @@ function readBoxVertical(
   return { ...recognizeQuad(recCtx, pageSrc, ordered), stacked: false };
 }
 
+// Pre-allocates one detector scratch-set per detect bucket (channel-independent
+// tensors; the per-run source-resize tensor is allocated in detectQuads). Mirrors
+// buildRecognizerSets — runs at construction.
+function buildDetectorSets(
+  detBuckets: readonly number[],
+  detectorKind: 'craft' | 'dbnet',
+  detExtraChannels: readonly number[]
+): DetSet[] {
+  return detBuckets.map((s) => {
+    const heat = s / 2;
+    return {
+      s,
+      tColor: tensor('uint8', [s, s, 3]),
+      tCF: tensor('uint8', [3, s, s]),
+      tNorm: tensor('float32', [3, s, s]),
+      tInput: tensor('float32', [1, 3, s, s]),
+      tHeatmap:
+        detectorKind === 'dbnet'
+          ? tensor('float32', [1, 1, s, s])
+          : tensor('float32', [1, heat, heat, 2]),
+      tExtras: detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat])),
+    };
+  });
+}
+
 // Pre-allocates one recognizer tensor-set per width bucket (each `recognize_<W>`
 // validated once) and derives the constant channel/height/vocab contract from the
 // first bucket. Kept out of the task factory; runs at construction.
@@ -734,6 +756,17 @@ export async function createOCR(
   let charset: string[] = [];
   let recSetByWidth: ReadonlyMap<number, RecSet> = new Map();
   let detExtraChannels: number[] = [];
+  let detSets: DetSet[] = [];
+  let detSetByS: ReadonlyMap<number, DetSet> = new Map();
+  const disposeDetSets = () =>
+    detSets.forEach((d) => {
+      d.tColor.dispose();
+      d.tCF.dispose();
+      d.tNorm.dispose();
+      d.tInput.dispose();
+      d.tHeatmap.dispose();
+      d.tExtras.forEach((t) => t.dispose());
+    });
   try {
     if (detBuckets.length === 0 || recBuckets.length === 0) {
       throw new Error(
@@ -781,6 +814,8 @@ export async function createOCR(
     }
     // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
     detExtraChannels = detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
+    detSets = buildDetectorSets(detBuckets, ocrOpts.detectorKind, detExtraChannels);
+    detSetByS = new Map(detSets.map((d) => [d.s, d]));
   } catch (e) {
     recSets.forEach((s) => {
       s.tCanvas.dispose();
@@ -789,6 +824,7 @@ export async function createOCR(
       s.tInput.dispose();
       s.tLogits.dispose();
     });
+    disposeDetSets();
     model.dispose();
     throw e;
   }
@@ -820,6 +856,7 @@ export async function createOCR(
       s.tInput.dispose();
       s.tLogits.dispose();
     });
+    disposeDetSets();
     model.dispose();
   };
 
@@ -851,13 +888,12 @@ export async function createOCR(
     // Detector state, reused for the page pass and the per-box character pass.
     const detCtx: DetectContext = {
       model,
-      detectorKind: ocrOpts.detectorKind,
       detBuckets,
       format,
       numChannels,
       detCode: rgbCode,
       extractOpts,
-      detExtraChannels,
+      detSets: detSetByS,
     };
 
     const tInputRaw = tensor('uint8', [height, width, numChannels]);

From 9c90617e0939e52fbfd30d3f04262c0b9b79e843 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:29:18 +0200
Subject: [PATCH 13/29] [RNE Rewrite] perf(cv): pre-decode boxes once in NMS
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The NMS suppression loop re-decoded box j via decodeToXyxy on every (i,j)
pair — O(N^2) decodes. Decode each candidate to xyxy+area once up front,
indexed by candidate position, and have both loops read the cached values.
Same result; decode work drops to O(N).
---
 .../cpp/extensions/cv/box_ops.cpp             | 57 +++++++++++--------
 1 file changed, 33 insertions(+), 24 deletions(-)

diff --git a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
index e1de7fccf6..031a5c948b 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
@@ -166,21 +166,37 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
         std::vector<std::vector<std::int32_t>> groups;
         std::vector<bool> suppressed(candidates.size(), false);
 
+        // Decode every candidate's box to xyxy + area once, indexed by candidate
+        // position, so the O(N^2) suppression loop below reads them instead of
+        // re-decoding box j on every pass.
+        struct DecodedBox {
+            float xmin, ymin, xmax, ymax, area;
+        };
+        std::vector<DecodedBox> decoded(candidates.size());
+        for (size_t k = 0; k < candidates.size(); ++k) {
+            const std::int32_t idx = candidates[k].first;
+            auto [xmin, ymin, xmax, ymax] = decodeToXyxy(
+                boxesPtr[idx * 4 + 0],
+                boxesPtr[idx * 4 + 1],
+                boxesPtr[idx * 4 + 2],
+                boxesPtr[idx * 4 + 3],
+                boxFormat);
+            decoded[k] = {
+                .xmin = xmin,
+                .ymin = ymin,
+                .xmax = xmax,
+                .ymax = ymax,
+                .area = (xmax - xmin) * (ymax - ymin)};
+        }
+
         for (size_t i = 0; i < candidates.size(); ++i) {
             if (suppressed[i]) {
                 continue;
             }
 
-            std::int32_t idxI = candidates[i].first;
-
-            auto [xminA, yminA, xmaxA, ymaxA] = decodeToXyxy(
-                boxesPtr[idxI * 4 + 0],
-                boxesPtr[idxI * 4 + 1],
-                boxesPtr[idxI * 4 + 2],
-                boxesPtr[idxI * 4 + 3],
-                boxFormat);
-
-            const float areaA = (xmaxA - xminA) * (ymaxA - yminA);
+            const std::int32_t idxI = candidates[i].first;
+            const DecodedBox &a = decoded[i];
+            const float areaA = a.area;
 
             std::vector<std::int32_t> overlapping = {idxI};
 
@@ -189,21 +205,14 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
                     continue;
                 }
 
-                std::int32_t idxJ = candidates[j].first;
-
-                auto [xminB, yminB, xmaxB, ymaxB] = decodeToXyxy(
-                    boxesPtr[idxJ * 4 + 0],
-                    boxesPtr[idxJ * 4 + 1],
-                    boxesPtr[idxJ * 4 + 2],
-                    boxesPtr[idxJ * 4 + 3],
-                    boxFormat);
-
-                const float areaB = (xmaxB - xminB) * (ymaxB - yminB);
+                const std::int32_t idxJ = candidates[j].first;
+                const DecodedBox &b = decoded[j];
+                const float areaB = b.area;
 
-                const float interYMin = std::max(yminA, yminB);
-                const float interXMin = std::max(xminA, xminB);
-                const float interYMax = std::min(ymaxA, ymaxB);
-                const float interXMax = std::min(xmaxA, xmaxB);
+                const float interYMin = std::max(a.ymin, b.ymin);
+                const float interXMin = std::max(a.xmin, b.xmin);
+                const float interYMax = std::min(a.ymax, b.ymax);
+                const float interXMax = std::min(a.xmax, b.xmax);
 
                 const float interH = std::max(0.0f, interYMax - interYMin);
                 const float interW = std::max(0.0f, interXMax - interXMin);

From 02e9faa091b296d39bfa7302849451f46ff5c3b9 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 15:45:28 +0200
Subject: [PATCH 14/29] [RNE Rewrite] feat(ocr): custom detector arch via TS
 box extraction
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

detectorKind gains 'custom': the model's raw detect_<S> outputs (shapes
read from the PTE method metadata, allocated for you) are handed to an
extractBoxes(outputs, s) worklet that returns quads in detector space —
the pipeline maps them to image pixels and applies dropScore, exactly
like the built-in craft/dbnet decoders. Pairs with the recognizer decode
hook so a fully foreign architecture slots in as config. Built-in paths
unchanged (DetSet now holds a tOutputs list; tOutputs[0] is the heatmap).
extractBoxes must be a worklet.
---
 .../src/extensions/cv/tasks/ocr.ts            | 180 ++++++++++++------
 1 file changed, 117 insertions(+), 63 deletions(-)

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 6f820094fc..06781d5312 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -20,6 +20,7 @@ import {
   type ColorConversionCode,
   type CraftExtractOptions,
   type DbnetExtractOptions,
+  type Quad,
 } from '../ops/image';
 import {
   buildCharset,
@@ -49,9 +50,11 @@ export type { Buckets } from './ocrHelpers';
 export type OCROptions = {
   /**
    * Detector architecture — selects the box decoder (CRAFT heatmap grouping vs
-   * DBNet prob-map) and the default drop score. A new architecture adds a variant.
+   * DBNet prob-map) and the default drop score. Use `'custom'` for any other
+   * architecture and supply {@link OCROptions.extractBoxes} to turn the raw
+   * detector output into quads in TypeScript.
    */
-  readonly detectorKind: 'craft' | 'dbnet';
+  readonly detectorKind: 'craft' | 'dbnet' | 'custom';
   /**
    * The model's static input-size buckets. The pipeline snaps each image to the
    * closest `detect`/`recognize` bucket and calls the matching per-size method
@@ -65,6 +68,16 @@ export type OCROptions = {
   readonly charset: string | readonly string[];
   /** Drop detections below this confidence. Defaults per detector architecture. */
   readonly dropScore?: number;
+  /**
+   * Custom detector post-processing, required when `detectorKind === 'custom'`.
+   * Receives the raw `detect_<S>` output tensors (the model's declared outputs,
+   * in order — shapes read from the PTE, allocated for you) and the snapped
+   * square side `s`, and returns oriented quads in DETECTOR space (the `s × s`
+   * letterboxed input); the pipeline maps them to image pixels and applies
+   * dropScore. Ignored for the built-in kinds. MUST be a worklet — it runs on
+   * the pipeline's worklet thread.
+   */
+  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
   /**
    * Recognizer input normalization, applied after the warp as `x·alpha + beta`
    * (scalar, or per-RGB-channel `[r,g,b]`). Defaults to `(x/255 − 0.5)/0.5` →
@@ -124,7 +137,11 @@ export type RunOCROptions = {
 // Everything else is shared — detector input is raw RGB /255 (mean/std baked in),
 // the recognizer is RGB with constant-128 left padding, both heads emit softmaxed
 // probabilities, and confidence is the mean of per-character max-probs.
-const DEFAULT_DROP_SCORE: Record<'craft' | 'dbnet', number> = { craft: 0, dbnet: 0.5 };
+const DEFAULT_DROP_SCORE: Record<'craft' | 'dbnet' | 'custom', number> = {
+  craft: 0,
+  dbnet: 0.5,
+  custom: 0,
+};
 
 /**
  * Model configuration required to instantiate an OCR task runner. One fused PTE
@@ -214,8 +231,9 @@ type DetSet = {
   readonly tCF: Tensor; // [3, s, s]
   readonly tNorm: Tensor; // [3, s, s]
   readonly tInput: Tensor; // [1, 3, s, s]
-  readonly tHeatmap: Tensor; // dbnet [1,1,s,s] | craft [1, s/2, s/2, 2]
-  readonly tExtras: readonly Tensor[]; // craft extras at half-res
+  // The detector's output tensors (dbnet prob-map / craft heatmap+extras / a
+  // custom arch's raw outputs); the built-in decoder reads tOutputs[0].
+  readonly tOutputs: readonly Tensor[];
 };
 
 // Everything the detector pass needs, bundled so it can run both on the full
@@ -226,7 +244,10 @@ type DetectContext = {
   readonly format: ImageFormat;
   readonly numChannels: number;
   readonly detCode: ColorConversionCode | null;
-  readonly extractOpts: DetectorExtractConfig;
+  // Built-in decode config (craft/dbnet); undefined when a custom extractor is used.
+  readonly extractOpts?: DetectorExtractConfig;
+  // Custom TS box extractor for detectorKind 'custom'; takes precedence when set.
+  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
   readonly detSets: ReadonlyMap<number, DetSet>;
 };
 
@@ -254,14 +275,18 @@ function detectQuads(
       .through(normalize, ds.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
       .copyTo(ds.tInput);
 
-    ctx.model.execute(`detect_${detS}`, [ds.tInput], [ds.tHeatmap, ...ds.tExtras]);
-    // CRAFT needs the per-run input height to restore its half-res boxes;
-    // `charLevel` switches it to per-glyph (ungrouped) boxes for a column pass.
-    const extractOpts =
-      ctx.extractOpts.mode === 'craft'
-        ? { ...ctx.extractOpts, targetHeight: detS, charLevel }
-        : ctx.extractOpts;
-    const quads = extractTextBoxes(ds.tHeatmap, extractOpts);
+    ctx.model.execute(`detect_${detS}`, [ds.tInput], [...ds.tOutputs]);
+    // A custom arch hands its raw outputs to the user extractor; the built-ins
+    // decode the heatmap (tOutputs[0]). CRAFT needs the per-run input height to
+    // restore its half-res boxes; `charLevel` switches to per-glyph boxes.
+    const quads = ctx.extractBoxes
+      ? ctx.extractBoxes(ds.tOutputs, detS)
+      : extractTextBoxes(
+          ds.tOutputs[0]!,
+          ctx.extractOpts!.mode === 'craft'
+            ? { ...ctx.extractOpts!, targetHeight: detS, charLevel }
+            : ctx.extractOpts!
+        );
     return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
   } finally {
     tDetResize.dispose();
@@ -653,23 +678,34 @@ function readBoxVertical(
 // tensors; the per-run source-resize tensor is allocated in detectQuads). Mirrors
 // buildRecognizerSets — runs at construction.
 function buildDetectorSets(
+  model: Model,
   detBuckets: readonly number[],
-  detectorKind: 'craft' | 'dbnet',
+  detectorKind: 'craft' | 'dbnet' | 'custom',
   detExtraChannels: readonly number[]
 ): DetSet[] {
   return detBuckets.map((s) => {
     const heat = s / 2;
+    // Custom archs declare arbitrary outputs — size them straight from the PTE's
+    // method metadata. Built-ins keep their known heatmap (+ craft extras) shapes.
+    let tOutputs: Tensor[];
+    if (detectorKind === 'custom') {
+      tOutputs = model
+        .getMethodMeta(`detect_${s}`)
+        .outputTensorMeta.map((m) => tensor(m.dtype, m.shape));
+    } else {
+      const tHeatmap =
+        detectorKind === 'dbnet'
+          ? tensor('float32', [1, 1, s, s])
+          : tensor('float32', [1, heat, heat, 2]);
+      tOutputs = [tHeatmap, ...detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]))];
+    }
     return {
       s,
       tColor: tensor('uint8', [s, s, 3]),
       tCF: tensor('uint8', [3, s, s]),
       tNorm: tensor('float32', [3, s, s]),
       tInput: tensor('float32', [1, 3, s, s]),
-      tHeatmap:
-        detectorKind === 'dbnet'
-          ? tensor('float32', [1, 1, s, s])
-          : tensor('float32', [1, heat, heat, 2]),
-      tExtras: detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat])),
+      tOutputs,
     };
   });
 }
@@ -764,8 +800,7 @@ export async function createOCR(
       d.tCF.dispose();
       d.tNorm.dispose();
       d.tInput.dispose();
-      d.tHeatmap.dispose();
-      d.tExtras.forEach((t) => t.dispose());
+      d.tOutputs.forEach((t) => t.dispose());
     });
   try {
     if (detBuckets.length === 0 || recBuckets.length === 0) {
@@ -773,28 +808,46 @@ export async function createOCR(
         'OCR: buckets.detect and buckets.recognize must each list at least one size.'
       );
     }
-    // Detector buckets feed a half-resolution CRAFT heatmap, so every side must be even.
-    if (detBuckets.some((s) => s % 2 !== 0)) {
-      throw new Error('OCR: every detect bucket side must be even (half-resolution heatmap).');
+    // CRAFT's half-resolution heatmap needs even detect-bucket sides.
+    if (ocrOpts.detectorKind === 'craft' && detBuckets.some((s) => s % 2 !== 0)) {
+      throw new Error(
+        'OCR: every CRAFT detect bucket side must be even (half-resolution heatmap).'
+      );
     }
-    // Validate every detect bucket (heatmap layout is constant across sizes); keep
-    // the largest bucket's meta for the constant extra-output channels.
     const detInSpec = [SymbolicTensor('float32', [1, 3, 'H', 'W'])];
-    const detOutSpec =
-      ocrOpts.detectorKind === 'dbnet'
-        ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
-        : [
-            SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
-            SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
-          ];
-    const detMeta = validateModelSchema(
-      model,
-      `detect_${detBuckets[detBuckets.length - 1]}`,
-      detInSpec,
-      detOutSpec
-    );
-    for (let i = 0; i < detBuckets.length - 1; i++) {
-      validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
+    if (ocrOpts.detectorKind === 'custom') {
+      if (!ocrOpts.extractBoxes) {
+        throw new Error("OCR: detectorKind 'custom' requires an extractBoxes worklet.");
+      }
+      // Outputs are arbitrary (read from metadata, handed to extractBoxes); only
+      // the shared RGB input contract is enforced. getMethodMeta throws if missing.
+      for (const s of detBuckets) {
+        const inShape = model.getMethodMeta(`detect_${s}`).inputTensorMeta[0]?.shape;
+        if (!inShape || inShape.length !== 4 || inShape[1] !== 3) {
+          throw new Error(`OCR: detect_${s} must take a [1, 3, ${s}, ${s}] RGB input.`);
+        }
+      }
+    } else {
+      // Validate every detect bucket against the architecture's output spec; keep
+      // the largest bucket's meta for the constant CRAFT extra-output channels.
+      const detOutSpec =
+        ocrOpts.detectorKind === 'dbnet'
+          ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
+          : [
+              SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
+              SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
+            ];
+      const detMeta = validateModelSchema(
+        model,
+        `detect_${detBuckets[detBuckets.length - 1]}`,
+        detInSpec,
+        detOutSpec
+      );
+      for (let i = 0; i < detBuckets.length - 1; i++) {
+        validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
+      }
+      // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
+      detExtraChannels = detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
     }
 
     const built = buildRecognizerSets(model, recBuckets);
@@ -812,9 +865,7 @@ export async function createOCR(
         `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${built.vocabSize}).`
       );
     }
-    // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
-    detExtraChannels = detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
-    detSets = buildDetectorSets(detBuckets, ocrOpts.detectorKind, detExtraChannels);
+    detSets = buildDetectorSets(model, detBuckets, ocrOpts.detectorKind, detExtraChannels);
     detSetByS = new Map(detSets.map((d) => [d.s, d]));
   } catch (e) {
     recSets.forEach((s) => {
@@ -829,24 +880,26 @@ export async function createOCR(
     throw e;
   }
 
-  // The extractTextBoxes mode matches detectorKind ('craft'/'dbnet').
-  const extractOpts: DetectorExtractConfig =
-    ocrOpts.detectorKind === 'dbnet'
-      ? {
-          mode: 'dbnet',
-          binThreshold: DBNET_BIN_THRESHOLD,
-          boxThreshold: DBNET_BOX_THRESHOLD,
-          unclipRatio: DBNET_UNCLIP_RATIO,
-          minBoxSide: DBNET_MIN_BOX_SIDE,
-          maxCandidates: DBNET_MAX_CANDIDATES,
-          applySigmoid: APPLY_SIGMOID,
-        }
-      : {
-          mode: 'craft',
-          textThreshold: CRAFT_TEXT_THRESHOLD,
-          linkThreshold: CRAFT_LINK_THRESHOLD,
-          lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
-        };
+  // Built-in box-decode config (custom archs decode in TS via extractBoxes instead).
+  const extractOpts: DetectorExtractConfig | undefined =
+    ocrOpts.detectorKind === 'custom'
+      ? undefined
+      : ocrOpts.detectorKind === 'dbnet'
+        ? {
+            mode: 'dbnet',
+            binThreshold: DBNET_BIN_THRESHOLD,
+            boxThreshold: DBNET_BOX_THRESHOLD,
+            unclipRatio: DBNET_UNCLIP_RATIO,
+            minBoxSide: DBNET_MIN_BOX_SIDE,
+            maxCandidates: DBNET_MAX_CANDIDATES,
+            applySigmoid: APPLY_SIGMOID,
+          }
+        : {
+            mode: 'craft',
+            textThreshold: CRAFT_TEXT_THRESHOLD,
+            linkThreshold: CRAFT_LINK_THRESHOLD,
+            lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
+          };
 
   const dispose = () => {
     recSets.forEach((s) => {
@@ -893,6 +946,7 @@ export async function createOCR(
       numChannels,
       detCode: rgbCode,
       extractOpts,
+      extractBoxes: ocrOpts.extractBoxes,
       detSets: detSetByS,
     };
 

From 603ad724672b10f6d920a3df674d561e42176697 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 16:21:26 +0200
Subject: [PATCH 15/29] [RNE Rewrite] docs(ocr): fix stale comments + naming
 after extensibility work

- points.ts: move scalePoint's JSDoc back onto scalePoint (the resizeFactors
  insertion had orphaned it).
- ocr.ts: update the baked-contract comments now that recognizer norm/pad/
  decode and the detector are per-model overridable; mark RECOGNIZER_* and the
  detectQuads scratch comments as defaults / cached.
- ocrHelpers.ts: rename the within-line sort helper cx -> xSum (it returns the
  edge sum, not the center; avoids clashing with the column-center cx).
- ImageViewport.tsx: boxes are in the displayed image's px, not 'original'.
---
 .../components/ImageViewport.tsx              |  8 ++---
 .../src/extensions/cv/ops/points.ts           | 24 +++++++--------
 .../src/extensions/cv/tasks/ocr.ts            | 29 ++++++++++++-------
 .../src/extensions/cv/tasks/ocrHelpers.ts     |  5 ++--
 4 files changed, 37 insertions(+), 29 deletions(-)

diff --git a/apps/computer-vision/components/ImageViewport.tsx b/apps/computer-vision/components/ImageViewport.tsx
index 701c83d655..e92684f43c 100644
--- a/apps/computer-vision/components/ImageViewport.tsx
+++ b/apps/computer-vision/components/ImageViewport.tsx
@@ -14,9 +14,9 @@ import { theme } from '../theme';
 const VIEW_WIDTH = Dimensions.get('window').width - 32;
 const DEFAULT_VIEW_HEIGHT = Math.round((VIEW_WIDTH * 16) / 9);
 
-/** A 2D point in original-image pixel coordinates. */
+/** A 2D point in the displayed image's pixel coordinates. */
 type Point = { readonly x: number; readonly y: number };
-/** A polygon (e.g. an OCR quad) in original-image pixel coordinates. */
+/** A polygon (e.g. an OCR quad) in the displayed image's pixel coordinates. */
 type Polygon = readonly Point[];
 
 export interface ImageViewportProps {
@@ -29,7 +29,7 @@ export interface ImageViewportProps {
   children?: React.ReactNode;
   /** Height of the preview box in px. Defaults to a 16:9 box. */
   height?: number;
-  /** Polygons (in original image px) to stroke over the image, e.g. OCR quads. */
+  /** Polygons (in the displayed image's px) to stroke over the image, e.g. OCR quads. */
   boxes?: readonly Polygon[];
 }
 
@@ -46,7 +46,7 @@ export function ImageViewport({
 }: ImageViewportProps) {
   const viewHeight = height ?? DEFAULT_VIEW_HEIGHT;
 
-  // Map original-pixel polygons into canvas space using the same contain-fit
+  // Map image-pixel polygons into canvas space using the same contain-fit
   // transform Skia uses to draw the image, then build one stroked path.
   const boxesPath = useMemo(() => {
     if (!skiaImage || !boxes?.length) return null;
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 524c2b7f25..e7d8f7c71d 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -22,18 +22,6 @@ export function clamp(v: number, lo: number, hi: number): number {
   return v < lo ? lo : v > hi ? hi : v;
 }
 
-/**
- * Helper function to scale a 2D point based on resize mode and resolution
- * changes.
- * @category Utils
- * @param point The original coordinate point to scale.
- * @param opts Options detailing the scaling factors and resize mode.
- * @param opts.from The source bounds (e.g. model input dimensions).
- * @param opts.to The destination bounds (e.g. original image dimensions).
- * @param opts.resizeMode The mode used to resize the image ('letterbox' or
- * 'stretch').
- * @returns The scaled coordinate point.
- */
 /**
  * Per-axis scale and offset that map a `to`-space coordinate back into
  * `from`-space, for an aspect-preserving letterbox or an axis stretch. The
@@ -63,6 +51,18 @@ export function resizeFactors(
   return { scaleX: from.width / to.width, scaleY: from.height / to.height, offsetX: 0, offsetY: 0 };
 }
 
+/**
+ * Helper function to scale a 2D point based on resize mode and resolution
+ * changes.
+ * @category Utils
+ * @param point The original coordinate point to scale.
+ * @param opts Options detailing the scaling factors and resize mode.
+ * @param opts.from The source bounds (e.g. model input dimensions).
+ * @param opts.to The destination bounds (e.g. original image dimensions).
+ * @param opts.resizeMode The mode used to resize the image ('letterbox' or
+ * 'stretch').
+ * @returns The scaled coordinate point.
+ */
 export function scalePoint(
   point: Point,
   opts: {
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 06781d5312..351939e6af 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -41,10 +41,12 @@ import {
 export type { Buckets } from './ocrHelpers';
 
 /**
- * Configuration for the unified OCR pipeline. The detector and recognizer share
- * one baked contract; only the box decoder (selected by `detectorKind`) and the
- * default drop score differ per architecture. A model declares its architecture,
- * its input-size buckets, and its charset.
+ * Configuration for the unified OCR pipeline. A model declares its detector
+ * architecture, its input-size buckets, and its charset; the detector/recognizer
+ * share one baked contract whose defaults match CRAFT (EasyOCR) and DBNet
+ * (PaddleOCR). Models that diverge can override the recognizer normalization,
+ * padding, and decode, or supply a `'custom'` detector with its own box
+ * extraction — see the per-field options below.
  * @category Types
  */
 export type OCROptions = {
@@ -132,11 +134,13 @@ export type RunOCROptions = {
   readonly release?: boolean;
 };
 
-// The unified baked contract leaves only two things per detector architecture:
-// the box decoder (selected by `detectorKind`) and the default drop score.
-// Everything else is shared — detector input is raw RGB /255 (mean/std baked in),
-// the recognizer is RGB with constant-128 left padding, both heads emit softmaxed
-// probabilities, and confidence is the mean of per-character max-probs.
+// Defaults for the shared baked contract — the detector input is raw RGB /255
+// (mean/std baked into the PTE), the recognizer is RGB with (x/255−0.5)/0.5 norm
+// and constant-128 left padding, both heads emit softmaxed probabilities, and
+// confidence is the mean of per-character max-probs. CRAFT/DBNet decode the
+// heatmap natively; everything else can be overridden per model via OCROptions
+// (recognizerNorm/recognizerPadValue/decode, and 'custom' detectorKind+extractBoxes).
+// Per-architecture default drop score:
 const DEFAULT_DROP_SCORE: Record<'craft' | 'dbnet' | 'custom', number> = {
   craft: 0,
   dbnet: 0.5,
@@ -177,9 +181,11 @@ export type OCRResult = {
   readonly detections: OCRDetection[];
 };
 
+// Default recognizer normalization / pad (SVTR/CRNN); overridable per model via
+// OCROptions.recognizerNorm / recognizerPadValue.
 const RECOGNIZER_ALPHA = 1 / 127.5; // (x/255 - 0.5)/0.5 -> [-1, 1]
 const RECOGNIZER_BETA = -1;
-const RECOGNIZER_PAD_VALUE = 128; // neutral gray; constant pad for both recognizers
+const RECOGNIZER_PAD_VALUE = 128; // neutral gray
 // Detector input is raw RGB scaled to [0,1]; the per-architecture mean/std is
 // baked into the PTE, so the client only divides by 255.
 const DETECTOR_ALPHA = 1 / 255;
@@ -253,7 +259,8 @@ type DetectContext = {
 
 // Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
 // returns quads in `src` pixel space: letterbox into the snapped square bucket,
-// run `detect_<S>`, extract quads. Scratch is per-size, allocated and freed here.
+// run `detect_<S>`, extract quads. Scratch is the bucket's cached DetSet; only
+// the source-resize tensor (input-channel-dependent) is allocated/freed here.
 function detectQuads(
   ctx: DetectContext,
   src: Tensor,
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
index 9a5ed12b37..82067739fd 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
@@ -281,9 +281,10 @@ export function readingOrderIndices(quads: readonly (readonly Point[])[]): numbe
       }
     }
     rows.sort((a, b) => a.ymin - b.ymin);
-    const cx = (i: number): number => boxes[i]!.xmin + boxes[i]!.xmax;
+    // Sort within a line by horizontal position (edge sum = 2× center; monotonic).
+    const xSum = (i: number): number => boxes[i]!.xmin + boxes[i]!.xmax;
     for (const row of rows) {
-      row.items.sort((a, b) => cx(a) - cx(b));
+      row.items.sort((a, b) => xSum(a) - xSum(b));
       for (const i of row.items) {
         out.push(i);
       }

From 6bd1d99f3529146452488a06a02c76e1bf5619c3 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 16:21:41 +0200
Subject: [PATCH 16/29] [RNE Rewrite] chore(ocr): include-what-you-use in
 ocr_ops.cpp

Include <jsi/jsi.h> directly (65 jsi:: uses, previously transitive) and
<opencv2/core/check.hpp>.
---
 packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 1d3fe6c685..10f3765af1 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -4,8 +4,10 @@
 #include <array>
 #include <cmath>
 #include <cstddef>
+#include <jsi/jsi.h>
 #include <limits>
 #include <numeric>
+#include <opencv2/core/check.hpp>
 #include <optional>
 #include <span>
 #include <stdexcept>

From 665e8ff26a9e39f327e3936d6db50c74747cb559 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 17:14:26 +0200
Subject: [PATCH 17/29] [RNE Rewrite] refactor(ocr): extract pipeline engine
 into ocrPipeline.ts

Split ocr.ts (1059 lines) so the task file holds only the public API +
createOCR factory:

- Move the tensor-pipeline engine (detectQuads, recognizeQuad,
  recognizeGlyphStrip, readStackedColumn, readBoxVertical), the per-bucket
  builders, and their context/set types into a new internal ocrPipeline.ts
  (imported only by ocr.ts; not re-exported from the package index).
- Extract validateDetectorSchema / buildExtractOpts / disposeDetSets /
  disposeRecSets, removing the duplicated recognizer/detector dispose loops.
- Hoist inline helpers to module scope: pushDetection (ocr.ts), lerp + xSum
  (ocrHelpers.ts).
- Drop the unused DetectContext.format field.

ocr.ts 1059 -> 459 lines. Behavior-preserving; verified on-device (Android):
detector localization + horizontal/vertical recognition unchanged.
---
 .../src/extensions/cv/tasks/ocr.ts            | 734 ++----------------
 .../src/extensions/cv/tasks/ocrHelpers.ts     | 147 +++-
 .../src/extensions/cv/tasks/ocrPipeline.ts    | 555 +++++++++++++
 3 files changed, 767 insertions(+), 669 deletions(-)
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts

diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 351939e6af..e417e5182f 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -1,42 +1,41 @@
 import type { WorkletRuntime } from 'react-native-worklets';
 
 import { tensor, type Tensor } from '../../../core/tensor';
-import { loadModel, type Model } from '../../../core/model';
-import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
+import { loadModel } from '../../../core/model';
 import { wrapAsync } from '../../../core/runtime';
 
-import type { ImageBuffer, ImageFormat } from '../image';
+import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
-import {
-  FORMAT_CONVERSION,
-  FORMAT_CHANNELS,
-  resize,
-  cvtColor,
-  toChannelsFirst,
-  normalize,
-  extractTextBoxes,
-  warpQuad,
-  ctcGreedyDecode,
-  type ColorConversionCode,
-  type CraftExtractOptions,
-  type DbnetExtractOptions,
-  type Quad,
-} from '../ops/image';
+import { FORMAT_CHANNELS, cvtColor, type Quad } from '../ops/image';
 import {
   buildCharset,
-  mapQuadToImage,
   orderQuad,
   quadSize,
-  contentWidthFor,
-  flattenQuad,
-  decodeGreedy,
-  ctcConfidence,
   nowMs,
-  snapDetectBucket,
-  snapRecognizeBucket,
   readingOrderIndices,
+  boundingQuadOf,
+  groupVerticalColumns,
   type Buckets,
 } from './ocrHelpers';
+import {
+  toRgbCode,
+  detectQuads,
+  recognizeQuad,
+  recognizeGlyphStrip,
+  readBoxVertical,
+  validateDetectorSchema,
+  buildExtractOpts,
+  buildDetectorSets,
+  buildRecognizerSets,
+  disposeDetSets,
+  disposeRecSets,
+  type DetSet,
+  type RecSet,
+  type DetectContext,
+  type RecContext,
+  type VerticalContext,
+  type DetectorExtractConfig,
+} from './ocrPipeline';
 
 export type { Buckets } from './ocrHelpers';
 
@@ -182,396 +181,34 @@ export type OCRResult = {
 };
 
 // Default recognizer normalization / pad (SVTR/CRNN); overridable per model via
-// OCROptions.recognizerNorm / recognizerPadValue.
+// OCROptions.recognizerNorm / recognizerPadValue. Detector-side norm and the
+// box-extraction tuning live with the engine in ocrPipeline.ts.
 const RECOGNIZER_ALPHA = 1 / 127.5; // (x/255 - 0.5)/0.5 -> [-1, 1]
 const RECOGNIZER_BETA = -1;
 const RECOGNIZER_PAD_VALUE = 128; // neutral gray
-// Detector input is raw RGB scaled to [0,1]; the per-architecture mean/std is
-// baked into the PTE, so the client only divides by 255.
-const DETECTOR_ALPHA = 1 / 255;
-const DETECTOR_BETA = 0;
-
-// Detector box-extraction tuning. These are stable across models and not worth
-// exposing as options — bake them in. (CRAFT = EasyOCR, DBNet = PaddleOCR.)
-const CRAFT_TEXT_THRESHOLD = 0.4;
-const CRAFT_LINK_THRESHOLD = 0.4;
-const CRAFT_LOW_TEXT_THRESHOLD = 0.7;
-const DBNET_BIN_THRESHOLD = 0.3;
-const DBNET_BOX_THRESHOLD = 0.6;
-const DBNET_UNCLIP_RATIO = 1.5;
-const DBNET_MIN_BOX_SIDE = 3;
-const DBNET_MAX_CANDIDATES = 1000;
-// DBNet exports a post-sigmoid prob map, so don't re-apply sigmoid (true only for
-// raw-logit heads).
-const APPLY_SIGMOID = false;
 // A box taller than this ratio is read as an upright stacked column.
 const TALL_CROP_RATIO = 1.5;
 // Per-page cap on stacked-column re-detection passes (each is detector-scale).
 const MAX_VERTICAL_REDETECTIONS = 8;
-// Wider than this × its height = a horizontal line, never a vertical-column glyph.
-const VERTICAL_GLYPH_ASPECT = 1.6;
-// A box joins a column when its x-span overlaps the column's by this fraction and
-// the y-gap is within this × its height (loose — signage letter spacing varies).
-const VERTICAL_X_OVERLAP = 0.25;
-const VERTICAL_Y_GAP = 2.5;
 // Vertical reads are lower-confidence and opt-in, so they skip the drop-score gate.
 const VERTICAL_DROP_SCORE = 0;
 
-// cvtColor code to bring a source image format to RGB (both the detector and the
-// recognizer operate on RGB), or null when it already is RGB.
-function toRgbCode(format: ImageFormat): ColorConversionCode | null {
-  'worklet';
-  return FORMAT_CONVERSION[format].rgb;
-}
-
-// Stored at construction; CRAFT's `targetHeight` is per-run, so detectQuads adds it.
-type DetectorExtractConfig = Omit<CraftExtractOptions, 'targetHeight'> | DbnetExtractOptions;
-
-// Per-detect-bucket scratch tensors, allocated once at construction (like RecSet)
-// and reused across the page pass and per-box re-detects. Only the source-resize
-// tensor depends on the run's input channel count, so detectQuads allocates that
-// one per call and reuses the rest from here.
-type DetSet = {
-  readonly s: number;
-  readonly tColor: Tensor; // [s, s, 3]
-  readonly tCF: Tensor; // [3, s, s]
-  readonly tNorm: Tensor; // [3, s, s]
-  readonly tInput: Tensor; // [1, 3, s, s]
-  // The detector's output tensors (dbnet prob-map / craft heatmap+extras / a
-  // custom arch's raw outputs); the built-in decoder reads tOutputs[0].
-  readonly tOutputs: readonly Tensor[];
-};
-
-// Everything the detector pass needs, bundled so it can run both on the full
-// page and (for vertical text) on a single box crop to find its characters.
-type DetectContext = {
-  readonly model: Model;
-  readonly detBuckets: readonly number[];
-  readonly format: ImageFormat;
-  readonly numChannels: number;
-  readonly detCode: ColorConversionCode | null;
-  // Built-in decode config (craft/dbnet); undefined when a custom extractor is used.
-  readonly extractOpts?: DetectorExtractConfig;
-  // Custom TS box extractor for detectorKind 'custom'; takes precedence when set.
-  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
-  readonly detSets: ReadonlyMap<number, DetSet>;
-};
-
-// Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
-// returns quads in `src` pixel space: letterbox into the snapped square bucket,
-// run `detect_<S>`, extract quads. Scratch is the bucket's cached DetSet; only
-// the source-resize tensor (input-channel-dependent) is allocated/freed here.
-function detectQuads(
-  ctx: DetectContext,
-  src: Tensor,
-  width: number,
-  height: number,
-  charLevel = false
-): Point[][] {
-  'worklet';
-  const detS = snapDetectBucket(width, height, ctx.detBuckets);
-  // snapDetectBucket always returns one of detBuckets, so the set exists.
-  const ds = ctx.detSets.get(detS)!;
-  // Only the source resize depends on the run's channel count; the rest is cached.
-  const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
-  try {
-    src
-      .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
-      .throughIf(ctx.detCode !== null, cvtColor, ds.tColor, ctx.detCode!)
-      .through(toChannelsFirst, ds.tCF)
-      .through(normalize, ds.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
-      .copyTo(ds.tInput);
-
-    ctx.model.execute(`detect_${detS}`, [ds.tInput], [...ds.tOutputs]);
-    // A custom arch hands its raw outputs to the user extractor; the built-ins
-    // decode the heatmap (tOutputs[0]). CRAFT needs the per-run input height to
-    // restore its half-res boxes; `charLevel` switches to per-glyph boxes.
-    const quads = ctx.extractBoxes
-      ? ctx.extractBoxes(ds.tOutputs, detS)
-      : extractTextBoxes(
-          ds.tOutputs[0]!,
-          ctx.extractOpts!.mode === 'craft'
-            ? { ...ctx.extractOpts!, targetHeight: detS, charLevel }
-            : ctx.extractOpts!
-        );
-    return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
-  } finally {
-    tDetResize.dispose();
-  }
-}
-
-// A recognizer width bucket's pre-allocated tensor-set (one per width in the
-// model's `recognize` buckets).
-type RecSet = {
-  readonly width: number;
-  readonly tCanvas: Tensor;
-  readonly tCF: Tensor;
-  readonly tNorm: Tensor;
-  readonly tInput: Tensor;
-  readonly tLogits: Tensor;
-};
-
-// Recognizer state for reading one quad. The source image is passed per call,
-// so the same context reads both the full page and a per-box crop.
-type RecContext = {
-  readonly model: Model;
-  readonly recSetByWidth: ReadonlyMap<number, RecSet>;
-  readonly recBuckets: readonly number[];
-  readonly recH: number;
-  readonly charset: string[];
-  // Per-model recognizer normalization / pad (resolved from OCROptions defaults).
-  readonly normAlpha: number | readonly number[];
-  readonly normBeta: number | readonly number[];
-  readonly padValue: number;
-  // Optional custom decode; falls back to greedy CTC when absent.
-  readonly decode?: (
-    logits: Tensor,
-    charset: readonly string[]
-  ) => { readonly text: string; readonly confidence: number };
-};
-
-// Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap content width to a
-// recognizer bucket, warp -> normalize -> execute -> greedy-CTC decode.
-function recognizeQuad(
-  ctx: RecContext,
-  src: Tensor,
-  corners: readonly Point[]
-): { text: string; conf: number } {
-  'worklet';
-  const cs = quadSize(corners);
-  const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
-  const desiredW = contentWidthFor(cs.width, cs.height, ctx.recH, maxRec);
-  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
-  // snapRecognizeBucket always returns one of recBuckets, so the set exists.
-  const rs = ctx.recSetByWidth.get(bucketW)!;
-  const contentWidth = Math.min(desiredW, bucketW);
-  warpQuad(src, rs.tCanvas, flattenQuad(corners), {
-    contentWidth,
-    align: 'left',
-    padMode: 'constant',
-    padValue: ctx.padValue,
-  });
-  rs.tCanvas
-    .through(toChannelsFirst, rs.tCF)
-    .through(normalize, rs.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
-    .copyTo(rs.tInput);
-  ctx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
-  // A custom decode (e.g. attention/AR head) takes the raw logits; otherwise
-  // greedy CTC. Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
-  if (ctx.decode) {
-    const r = ctx.decode(rs.tLogits, ctx.charset);
-    return { text: r.text, conf: r.confidence };
-  }
-  const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
-  const text = decodeGreedy(indices, ctx.charset);
-  const conf = ctcConfidence(values, indices);
-  return { text, conf };
-}
-
-// State the vertical-text path needs on top of `RecContext`: the detector (for
-// the second, character-level pass) and the page tensors it crops the box from.
-type VerticalContext = {
-  readonly detCtx: DetectContext;
-  readonly rawPage: Tensor; // full page, native format — cropped per box for re-detection
-  readonly recCode: ColorConversionCode | null; // native format -> recognizer color
-  readonly recC: number;
-  readonly tallCropRatio: number; // height/width ratio above which a box is a column
-  // Per-page budget for the (expensive) stacked-column re-detection pass.
-  readonly redetectBudget: { remaining: number };
-};
-
-// Divides an ordered TL,TR,BR,BL box into `parts` equal vertical bands (each a
-// TL,TR,BR,BL quad), top -> bottom. Used to recover the individual upright
-// letters of a stacked column from a box the detector merged (DBNet emits one
-// box per text region, not per glyph, so stacked letters arrive fused). `parts`
-// <= 1 returns the box unchanged.
-function splitTallQuad(ordered: readonly Point[], parts: number): Point[][] {
-  'worklet';
-  if (parts <= 1) {
-    return [ordered as Point[]];
-  }
-  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
-  const lerp = (a: Point, b: Point, t: number): Point => ({
-    x: a.x + (b.x - a.x) * t,
-    y: a.y + (b.y - a.y) * t,
-  });
-  const out: Point[][] = [];
-  for (let i = 0; i < parts; i++) {
-    const t0 = i / parts;
-    const t1 = (i + 1) / parts;
-    // Left edge runs tl->bl, right edge tr->br; take the band between t0 and t1.
-    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
-  }
-  return out;
-}
-
-// Joins glyph quads (in `src` pixel space, in reading order) into one recognizer
-// strip — each glyph warped upright to the recognizer height and laid side by
-// side — and recognizes it as a single line (joint hconcat). Returns null when
-// nothing usable was assembled.
-//
-// Must be defined BEFORE its callers: the worklet plugin captures referenced
-// worklets in source order, so a forward reference is undefined at run time.
-function recognizeGlyphStrip(
-  recCtx: RecContext,
-  src: Tensor,
-  recC: number,
-  glyphs: readonly (readonly Point[])[]
-): { text: string; conf: number } | null {
-  'worklet';
-  const recH = recCtx.recH;
-  const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-  // Warp each glyph upright to recognizer height (aspect preserved). A box that
-  // is much taller than wide is a merged run of stacked letters — split it into
-  // ~square single-letter cells first, so each lands in its own strip slot
-  // (otherwise N letters get squashed into one cell and read as garbage).
-  const slices: { tGlyph: Tensor; w: number }[] = [];
-  let totalW = 0;
-  for (const g of glyphs) {
-    const gsz = quadSize(g);
-    if (gsz.width < 1 || gsz.height < 1) {
-      continue;
-    }
-    const parts = Math.max(1, Math.round(gsz.height / Math.max(1, gsz.width)));
-    for (const cell of splitTallQuad(g, parts)) {
-      const gs = quadSize(cell);
-      if (gs.width < 1 || gs.height < 1) {
-        continue;
-      }
-      const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
-      const tGlyph = tensor('uint8', [recH, gw, recC]);
-      warpQuad(src, tGlyph, flattenQuad(cell), {
-        contentWidth: gw,
-        align: 'left',
-        padMode: 'constant',
-        padValue: recCtx.padValue,
-      });
-      slices.push({ tGlyph, w: gw });
-      totalW += gw;
-    }
-  }
-  if (slices.length === 0) {
-    return null;
-  }
-  try {
-    // Smallest bucket that fits the strip (snap up, no glyph truncated); widest
-    // bucket for very long columns.
-    const bucketW =
-      recCtx.recBuckets.find((w) => w >= totalW) ??
-      recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-    const rs = recCtx.recSetByWidth.get(bucketW)!;
-    // Assemble the strip row-major into the bucket canvas, neutral-padded.
-    const strip = new Uint8Array(recH * bucketW * recC);
-    strip.fill(recCtx.padValue);
-    let xOff = 0;
-    for (const s of slices) {
-      if (xOff >= bucketW) {
-        break;
-      }
-      const copyW = Math.min(s.w, bucketW - xOff);
-      const glyphBytes = new Uint8Array(recH * s.w * recC);
-      s.tGlyph.getData(glyphBytes);
-      for (let oy = 0; oy < recH; oy++) {
-        const srcStart = oy * s.w * recC;
-        const row = glyphBytes.subarray(srcStart, srcStart + copyW * recC);
-        strip.set(row, (oy * bucketW + xOff) * recC);
-      }
-      xOff += s.w;
-    }
-    rs.tCanvas.setData(strip);
-    rs.tCanvas
-      .through(toChannelsFirst, rs.tCF)
-      .through(normalize, rs.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
-      .copyTo(rs.tInput);
-    recCtx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
-    if (recCtx.decode) {
-      const r = recCtx.decode(rs.tLogits, recCtx.charset);
-      return r.text.length > 0 ? { text: r.text, conf: r.confidence } : null;
-    }
-    const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
-    const text = decodeGreedy(indices, recCtx.charset);
-    const conf = ctcConfidence(values, indices);
-    return text.length > 0 ? { text, conf } : null;
-  } finally {
-    slices.forEach((s) => s.tGlyph.dispose());
-  }
-}
-
-// Reads a single tall box that packs several stacked glyphs the detector grouped
-// into one box (e.g. a tightly-set container code): crop it upright, re-detect
-// the glyphs (char-level pass), and read them top -> bottom as a joined strip.
-// Returns null — caller falls back to a horizontal read — when the box is tiny,
-// the per-page re-detect budget is spent, or nothing is found.
-function readStackedColumn(
-  recCtx: RecContext,
-  vctx: VerticalContext,
-  ordered: readonly Point[],
-  size: { width: number; height: number }
-): { text: string; conf: number } | null {
+// Appends a detection when it has text and clears the drop-score threshold. A
+// module-level worklet (not a closure) so the run loop stays flat.
+function pushDetection(
+  out: OCRDetection[],
+  threshold: number,
+  text: string,
+  conf: number,
+  quad: readonly Point[],
+  ms: number
+): void {
   'worklet';
-  const boxW = Math.round(size.width);
-  const boxH = Math.round(size.height);
-  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
-    return null;
-  }
-  vctx.redetectBudget.remaining--;
-  const recC = vctx.recC;
-  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
-  const tRecBox = tensor('uint8', [boxH, boxW, recC]);
-  try {
-    // Axis-aligned upright crop of the box, full width (no padding).
-    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
-      contentWidth: boxW,
-      align: 'left',
-      padMode: 'constant',
-      padValue: 0,
-    });
-    // Char-level second pass: per-glyph boxes (no grouping), in crop pixel space.
-    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
-    if (charQuads.length === 0) {
-      return null;
-    }
-    const boxSrc = vctx.recCode !== null ? cvtColor(tBoxRaw, tRecBox, vctx.recCode) : tBoxRaw;
-    // Stack reading order: top -> bottom by each glyph's upper edge.
-    const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
-    return recognizeGlyphStrip(recCtx, boxSrc, recC, glyphs);
-  } finally {
-    tBoxRaw.dispose();
-    tRecBox.dispose();
+  if (text.length > 0 && conf >= threshold) {
+    out.push({ text, confidence: conf, quad, recognizeMs: ms });
   }
 }
 
-// Bounding axis-aligned quad (TL,TR,BR,BL) enclosing a set of quads.
-function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
-  'worklet';
-  if (quads.length === 0) {
-    return [
-      { x: 0, y: 0 },
-      { x: 0, y: 0 },
-      { x: 0, y: 0 },
-      { x: 0, y: 0 },
-    ];
-  }
-  let xmin = Infinity;
-  let ymin = Infinity;
-  let xmax = -Infinity;
-  let ymax = -Infinity;
-  for (const q of quads) {
-    for (const p of q) {
-      xmin = Math.min(xmin, p.x);
-      ymin = Math.min(ymin, p.y);
-      xmax = Math.max(xmax, p.x);
-      ymax = Math.max(ymax, p.y);
-    }
-  }
-  return [
-    { x: xmin, y: ymin },
-    { x: xmax, y: ymin },
-    { x: xmax, y: ymax },
-    { x: xmin, y: ymax },
-  ];
-}
-
 // Reorders recognized detections into human reading order (the detector emits
 // boxes in an arbitrary order). Column-aware: genuine multi-column pages read
 // column-by-column, single-column pages line-by-line, words within a line
@@ -585,174 +222,6 @@ function orderDetections(dets: OCRDetection[]): OCRDetection[] {
   return order.map((i) => dets[i]!);
 }
 
-// Clusters glyph-like, x-aligned, stacked boxes into vertical columns; wide lines
-// and isolated boxes come back as `singles` to read normally. So `vertical` ADDS
-// column reading without disturbing horizontal reads.
-function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
-  columns: Point[][][];
-  singles: Point[][];
-} {
-  'worklet';
-  type B = {
-    q: Point[];
-    xmin: number;
-    xmax: number;
-    ymin: number;
-    ymax: number;
-    w: number;
-    h: number;
-  };
-  const candidates: B[] = [];
-  const singles: Point[][] = [];
-  for (const q of quads) {
-    let xmin = Infinity;
-    let ymin = Infinity;
-    let xmax = -Infinity;
-    let ymax = -Infinity;
-    for (const p of q) {
-      xmin = Math.min(xmin, p.x);
-      ymin = Math.min(ymin, p.y);
-      xmax = Math.max(xmax, p.x);
-      ymax = Math.max(ymax, p.y);
-    }
-    const w = xmax - xmin;
-    const h = ymax - ymin;
-    if (w > h * VERTICAL_GLYPH_ASPECT) {
-      singles.push(q as Point[]); // a horizontal line — read normally
-    } else {
-      candidates.push({ q: q as Point[], xmin, xmax, ymin, ymax, w, h });
-    }
-  }
-  // Top -> bottom, growing each column from its current bottom box. Alignment is
-  // checked against the column's accumulated x-range (not just the last box), so a
-  // narrow glyph like `I` between wider ones doesn't break the run.
-  candidates.sort((a, b) => a.ymin - b.ymin);
-  type Col = { boxes: B[]; xmin: number; xmax: number; bottom: number };
-  const cols: Col[] = [];
-  for (const b of candidates) {
-    let placed = false;
-    for (const col of cols) {
-      const overlap = Math.min(b.xmax, col.xmax) - Math.max(b.xmin, col.xmin);
-      const aligned = overlap > VERTICAL_X_OVERLAP * Math.min(b.w, col.xmax - col.xmin);
-      const gap = b.ymin - col.bottom;
-      if (aligned && gap < VERTICAL_Y_GAP * b.h && gap > -0.5 * b.h) {
-        col.boxes.push(b);
-        col.xmin = Math.min(col.xmin, b.xmin);
-        col.xmax = Math.max(col.xmax, b.xmax);
-        col.bottom = b.ymax;
-        placed = true;
-        break;
-      }
-    }
-    if (!placed) {
-      cols.push({ boxes: [b], xmin: b.xmin, xmax: b.xmax, bottom: b.ymax });
-    }
-  }
-  const columns: Point[][][] = [];
-  for (const col of cols) {
-    if (col.boxes.length >= 2) {
-      columns.push(col.boxes.map((b) => b.q)); // already top -> bottom
-    } else {
-      singles.push(col.boxes[0]!.q);
-    }
-  }
-  return { columns, singles };
-}
-
-// Reads one box that may be an upright stacked column (e.g. a shipping-container
-// code, letters stacked top-to-bottom). A normal-aspect box is read horizontally;
-// a tall box is read as a stacked column, falling back to the horizontal read.
-// `stacked` reports whether the column path produced the read (caller applies the
-// lenient vertical drop-score to those).
-function readBoxVertical(
-  recCtx: RecContext,
-  vctx: VerticalContext,
-  pageSrc: Tensor,
-  ordered: readonly Point[],
-  size: { width: number; height: number }
-): { text: string; conf: number; stacked: boolean } {
-  'worklet';
-  if (size.height >= size.width * vctx.tallCropRatio) {
-    const stacked = readStackedColumn(recCtx, vctx, ordered, size);
-    if (stacked) {
-      return { ...stacked, stacked: true };
-    }
-  }
-  return { ...recognizeQuad(recCtx, pageSrc, ordered), stacked: false };
-}
-
-// Pre-allocates one detector scratch-set per detect bucket (channel-independent
-// tensors; the per-run source-resize tensor is allocated in detectQuads). Mirrors
-// buildRecognizerSets — runs at construction.
-function buildDetectorSets(
-  model: Model,
-  detBuckets: readonly number[],
-  detectorKind: 'craft' | 'dbnet' | 'custom',
-  detExtraChannels: readonly number[]
-): DetSet[] {
-  return detBuckets.map((s) => {
-    const heat = s / 2;
-    // Custom archs declare arbitrary outputs — size them straight from the PTE's
-    // method metadata. Built-ins keep their known heatmap (+ craft extras) shapes.
-    let tOutputs: Tensor[];
-    if (detectorKind === 'custom') {
-      tOutputs = model
-        .getMethodMeta(`detect_${s}`)
-        .outputTensorMeta.map((m) => tensor(m.dtype, m.shape));
-    } else {
-      const tHeatmap =
-        detectorKind === 'dbnet'
-          ? tensor('float32', [1, 1, s, s])
-          : tensor('float32', [1, heat, heat, 2]);
-      tOutputs = [tHeatmap, ...detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]))];
-    }
-    return {
-      s,
-      tColor: tensor('uint8', [s, s, 3]),
-      tCF: tensor('uint8', [3, s, s]),
-      tNorm: tensor('float32', [3, s, s]),
-      tInput: tensor('float32', [1, 3, s, s]),
-      tOutputs,
-    };
-  });
-}
-
-// Pre-allocates one recognizer tensor-set per width bucket (each `recognize_<W>`
-// validated once) and derives the constant channel/height/vocab contract from the
-// first bucket. Kept out of the task factory; runs at construction.
-function buildRecognizerSets(
-  model: Model,
-  recBuckets: readonly number[]
-): { sets: RecSet[]; recC: number; recH: number; vocabSize: number } {
-  let recC = 0;
-  let recH = 0;
-  let vocabSize = 0;
-  const sets = recBuckets.map((w, i) => {
-    const m = validateModelSchema(
-      model,
-      `recognize_${w}`,
-      [SymbolicTensor('float32', [1, 'C', 'H', 'W'])],
-      [SymbolicTensor('float32', [1, 'T', 'V'])]
-    );
-    const inShape = m.inputTensorMeta[0]!.shape;
-    if (i === 0) {
-      // Channels/height/vocab are constant across the width buckets.
-      recC = inShape[1]!;
-      recH = inShape[2]!;
-      vocabSize = m.outputTensorMeta[0]!.shape[2]!;
-    }
-    return {
-      width: w,
-      tCanvas: tensor('uint8', [recH, w, recC]),
-      tCF: tensor('uint8', [recC, recH, w]),
-      tNorm: tensor('float32', [recC, recH, w]),
-      tInput: tensor('float32', inShape),
-      tLogits: tensor('float32', m.outputTensorMeta[0]!.shape),
-    };
-  });
-  return { sets, recC, recH, vocabSize };
-}
-
 /**
  * Creates a unified OCR runner for two-stage detect -> recognize models
  * (EasyOCR / PaddleOCR). It loads one fused PTE, validates the `detect` and
@@ -798,64 +267,20 @@ export async function createOCR(
   let recH = 0;
   let charset: string[] = [];
   let recSetByWidth: ReadonlyMap<number, RecSet> = new Map();
-  let detExtraChannels: number[] = [];
   let detSets: DetSet[] = [];
   let detSetByS: ReadonlyMap<number, DetSet> = new Map();
-  const disposeDetSets = () =>
-    detSets.forEach((d) => {
-      d.tColor.dispose();
-      d.tCF.dispose();
-      d.tNorm.dispose();
-      d.tInput.dispose();
-      d.tOutputs.forEach((t) => t.dispose());
-    });
   try {
     if (detBuckets.length === 0 || recBuckets.length === 0) {
       throw new Error(
         'OCR: buckets.detect and buckets.recognize must each list at least one size.'
       );
     }
-    // CRAFT's half-resolution heatmap needs even detect-bucket sides.
-    if (ocrOpts.detectorKind === 'craft' && detBuckets.some((s) => s % 2 !== 0)) {
-      throw new Error(
-        'OCR: every CRAFT detect bucket side must be even (half-resolution heatmap).'
-      );
-    }
-    const detInSpec = [SymbolicTensor('float32', [1, 3, 'H', 'W'])];
-    if (ocrOpts.detectorKind === 'custom') {
-      if (!ocrOpts.extractBoxes) {
-        throw new Error("OCR: detectorKind 'custom' requires an extractBoxes worklet.");
-      }
-      // Outputs are arbitrary (read from metadata, handed to extractBoxes); only
-      // the shared RGB input contract is enforced. getMethodMeta throws if missing.
-      for (const s of detBuckets) {
-        const inShape = model.getMethodMeta(`detect_${s}`).inputTensorMeta[0]?.shape;
-        if (!inShape || inShape.length !== 4 || inShape[1] !== 3) {
-          throw new Error(`OCR: detect_${s} must take a [1, 3, ${s}, ${s}] RGB input.`);
-        }
-      }
-    } else {
-      // Validate every detect bucket against the architecture's output spec; keep
-      // the largest bucket's meta for the constant CRAFT extra-output channels.
-      const detOutSpec =
-        ocrOpts.detectorKind === 'dbnet'
-          ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
-          : [
-              SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
-              SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
-            ];
-      const detMeta = validateModelSchema(
-        model,
-        `detect_${detBuckets[detBuckets.length - 1]}`,
-        detInSpec,
-        detOutSpec
-      );
-      for (let i = 0; i < detBuckets.length - 1; i++) {
-        validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
-      }
-      // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
-      detExtraChannels = detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
-    }
+    const detExtraChannels = validateDetectorSchema(
+      model,
+      detBuckets,
+      ocrOpts.detectorKind,
+      ocrOpts.extractBoxes
+    );
 
     const built = buildRecognizerSets(model, recBuckets);
     recSets = built.sets;
@@ -875,48 +300,18 @@ export async function createOCR(
     detSets = buildDetectorSets(model, detBuckets, ocrOpts.detectorKind, detExtraChannels);
     detSetByS = new Map(detSets.map((d) => [d.s, d]));
   } catch (e) {
-    recSets.forEach((s) => {
-      s.tCanvas.dispose();
-      s.tCF.dispose();
-      s.tNorm.dispose();
-      s.tInput.dispose();
-      s.tLogits.dispose();
-    });
-    disposeDetSets();
+    disposeRecSets(recSets);
+    disposeDetSets(detSets);
     model.dispose();
     throw e;
   }
 
-  // Built-in box-decode config (custom archs decode in TS via extractBoxes instead).
-  const extractOpts: DetectorExtractConfig | undefined =
-    ocrOpts.detectorKind === 'custom'
-      ? undefined
-      : ocrOpts.detectorKind === 'dbnet'
-        ? {
-            mode: 'dbnet',
-            binThreshold: DBNET_BIN_THRESHOLD,
-            boxThreshold: DBNET_BOX_THRESHOLD,
-            unclipRatio: DBNET_UNCLIP_RATIO,
-            minBoxSide: DBNET_MIN_BOX_SIDE,
-            maxCandidates: DBNET_MAX_CANDIDATES,
-            applySigmoid: APPLY_SIGMOID,
-          }
-        : {
-            mode: 'craft',
-            textThreshold: CRAFT_TEXT_THRESHOLD,
-            linkThreshold: CRAFT_LINK_THRESHOLD,
-            lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
-          };
+  // Built-in box-decode config (custom archs decode in TS via extractBoxes).
+  const extractOpts: DetectorExtractConfig | undefined = buildExtractOpts(ocrOpts.detectorKind);
 
   const dispose = () => {
-    recSets.forEach((s) => {
-      s.tCanvas.dispose();
-      s.tCF.dispose();
-      s.tNorm.dispose();
-      s.tInput.dispose();
-      s.tLogits.dispose();
-    });
-    disposeDetSets();
+    disposeRecSets(recSets);
+    disposeDetSets(detSets);
     model.dispose();
   };
 
@@ -949,7 +344,6 @@ export async function createOCR(
     const detCtx: DetectContext = {
       model,
       detBuckets,
-      format,
       numChannels,
       detCode: rgbCode,
       extractOpts,
@@ -993,14 +387,6 @@ export async function createOCR(
       };
 
       const detections: OCRDetection[] = [];
-      const pushAt =
-        (threshold: number) => (text: string, conf: number, quad: readonly Point[], ms: number) => {
-          if (text.length > 0 && conf >= threshold) {
-            detections.push({ text, confidence: conf, quad, recognizeMs: ms });
-          }
-        };
-      const push = pushAt(dropScore); // flat lines: normal drop-score
-      const pushVertical = pushAt(VERTICAL_DROP_SCORE); // stacked/column: lenient
 
       // Valid (non-tiny) boxes, ordered TL,TR,BR,BL.
       const ordered: Point[][] = [];
@@ -1016,7 +402,7 @@ export async function createOCR(
         for (const o of ordered) {
           const boxStart = nowMs();
           const { text, conf } = recognizeQuad(recCtx, recSrc, o);
-          push(text, conf, o, nowMs() - boxStart);
+          pushDetection(detections, dropScore, text, conf, o, nowMs() - boxStart);
         }
         return { detections: orderDetections(detections) };
       }
@@ -1031,13 +417,27 @@ export async function createOCR(
         // column's boxes can be passed straight through, top -> bottom.
         const r = recognizeGlyphStrip(recCtx, recSrc, recC, col);
         if (r) {
-          pushVertical(r.text, r.conf, boundingQuadOf(col), nowMs() - boxStart);
+          pushDetection(
+            detections,
+            VERTICAL_DROP_SCORE,
+            r.text,
+            r.conf,
+            boundingQuadOf(col),
+            nowMs() - boxStart
+          );
         }
       }
       for (const o of singles) {
         const boxStart = nowMs();
         const { text, conf, stacked } = readBoxVertical(recCtx, vctx, recSrc, o, quadSize(o));
-        (stacked ? pushVertical : push)(text, conf, o, nowMs() - boxStart);
+        pushDetection(
+          detections,
+          stacked ? VERTICAL_DROP_SCORE : dropScore,
+          text,
+          conf,
+          o,
+          nowMs() - boxStart
+        );
       }
       return { detections: orderDetections(detections) };
     } finally {
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
index 82067739fd..32555903c1 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
@@ -148,6 +148,13 @@ const distance = (a: Point, b: Point): number => {
   return Math.hypot(b.x - a.x, b.y - a.y);
 };
 
+// Linear interpolation between two points (t in [0,1]). Module-level so the
+// worklet plugin captures it; used by splitTallQuad to cut a quad into bands.
+const lerp = (a: Point, b: Point, t: number): Point => {
+  'worklet';
+  return { x: a.x + (b.x - a.x) * t, y: a.y + (b.y - a.y) * t };
+};
+
 /**
  * The natural width/height (in pixels) of an ordered TL,TR,BR,BL quad.
  * @category Typescript API
@@ -191,6 +198,13 @@ function bboxOfQuad(quad: readonly Point[]): ReadingBox {
   return { xmin, ymin, xmax, ymax };
 }
 
+// Horizontal sort key for a box: edge sum = 2× center-x (monotonic in center,
+// avoids the divide). Module-level so it isn't re-created per readingOrderIndices.
+function xSum(b: ReadingBox): number {
+  'worklet';
+  return b.xmin + b.xmax;
+}
+
 /**
  * Returns the indices of `quads` in human reading order. Detects vertical column
  * gutters with an x-coverage sweep — a band crossed by no box and wider than
@@ -282,9 +296,8 @@ export function readingOrderIndices(quads: readonly (readonly Point[])[]): numbe
     }
     rows.sort((a, b) => a.ymin - b.ymin);
     // Sort within a line by horizontal position (edge sum = 2× center; monotonic).
-    const xSum = (i: number): number => boxes[i]!.xmin + boxes[i]!.xmax;
     for (const row of rows) {
-      row.items.sort((a, b) => xSum(a) - xSum(b));
+      row.items.sort((a, b) => xSum(boxes[a]!) - xSum(boxes[b]!));
       for (const i of row.items) {
         out.push(i);
       }
@@ -293,6 +306,136 @@ export function readingOrderIndices(quads: readonly (readonly Point[])[]): numbe
   return out;
 }
 
+// Vertical-text grouping tuning. A box wider than this × its height is a
+// horizontal line, never a stacked-column glyph. A box joins a column when its
+// x-span overlaps the column's by VERTICAL_X_OVERLAP of the narrower width and the
+// y-gap is within VERTICAL_Y_GAP × its height (loose — signage spacing varies).
+const VERTICAL_GLYPH_ASPECT = 1.6;
+const VERTICAL_X_OVERLAP = 0.25;
+const VERTICAL_Y_GAP = 2.5;
+
+/**
+ * Divides an ordered TL,TR,BR,BL quad into `parts` equal vertical bands (each a
+ * TL,TR,BR,BL quad), top to bottom. Recovers the individual upright letters of a
+ * stacked column from a box the detector merged (DBNet emits one box per text
+ * region, not per glyph, so stacked letters arrive fused). `parts <= 1` returns
+ * the quad unchanged.
+ * @category Typescript API
+ * @param ordered The quad corners ordered TL, TR, BR, BL.
+ * @param parts The number of equal vertical bands to split into.
+ * @returns The bands as ordered TL,TR,BR,BL quads, top to bottom.
+ */
+export function splitTallQuad(ordered: readonly Point[], parts: number): Point[][] {
+  'worklet';
+  if (parts <= 1) {
+    return [ordered as Point[]];
+  }
+  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
+  const out: Point[][] = [];
+  for (let i = 0; i < parts; i++) {
+    const t0 = i / parts;
+    const t1 = (i + 1) / parts;
+    // Left edge runs tl->bl, right edge tr->br; take the band between t0 and t1.
+    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
+  }
+  return out;
+}
+
+/**
+ * The axis-aligned bounding quad (TL,TR,BR,BL) enclosing a set of quads. Returns a
+ * zero quad for empty input.
+ * @category Typescript API
+ * @param quads The quads to enclose.
+ * @returns The four corners of the enclosing box, ordered TL, TR, BR, BL.
+ */
+export function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
+  'worklet';
+  const all: Point[] = [];
+  for (const q of quads) {
+    for (const p of q) {
+      all.push(p);
+    }
+  }
+  const { xmin, ymin, xmax, ymax } = bboxOfQuad(all);
+  return [
+    { x: xmin, y: ymin },
+    { x: xmax, y: ymin },
+    { x: xmax, y: ymax },
+    { x: xmin, y: ymax },
+  ];
+}
+
+/**
+ * Clusters glyph-like, x-aligned, stacked boxes into vertical columns; wide lines
+ * and isolated boxes come back as `singles` to read normally. So the vertical pass
+ * ADDS column reading without disturbing horizontal reads.
+ * @category Typescript API
+ * @param quads The detected text quads (ordered TL,TR,BR,BL).
+ * @returns The detected `columns` (each a top-to-bottom list of quads) and the
+ * leftover `singles` (horizontal lines / isolated boxes) to read normally.
+ */
+export function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
+  columns: Point[][][];
+  singles: Point[][];
+} {
+  'worklet';
+  type B = {
+    q: Point[];
+    xmin: number;
+    xmax: number;
+    ymin: number;
+    ymax: number;
+    w: number;
+    h: number;
+  };
+  const candidates: B[] = [];
+  const singles: Point[][] = [];
+  for (const q of quads) {
+    const { xmin, ymin, xmax, ymax } = bboxOfQuad(q);
+    const w = xmax - xmin;
+    const h = ymax - ymin;
+    if (w > h * VERTICAL_GLYPH_ASPECT) {
+      singles.push(q as Point[]); // a horizontal line — read normally
+    } else {
+      candidates.push({ q: q as Point[], xmin, xmax, ymin, ymax, w, h });
+    }
+  }
+  // Top -> bottom, growing each column from its current bottom box. Alignment is
+  // checked against the column's accumulated x-range (not just the last box), so a
+  // narrow glyph like `I` between wider ones doesn't break the run.
+  candidates.sort((a, b) => a.ymin - b.ymin);
+  type Col = { boxes: B[]; xmin: number; xmax: number; bottom: number };
+  const cols: Col[] = [];
+  for (const b of candidates) {
+    let placed = false;
+    for (const col of cols) {
+      const overlap = Math.min(b.xmax, col.xmax) - Math.max(b.xmin, col.xmin);
+      const aligned = overlap > VERTICAL_X_OVERLAP * Math.min(b.w, col.xmax - col.xmin);
+      const gap = b.ymin - col.bottom;
+      if (aligned && gap < VERTICAL_Y_GAP * b.h && gap > -0.5 * b.h) {
+        col.boxes.push(b);
+        col.xmin = Math.min(col.xmin, b.xmin);
+        col.xmax = Math.max(col.xmax, b.xmax);
+        col.bottom = b.ymax;
+        placed = true;
+        break;
+      }
+    }
+    if (!placed) {
+      cols.push({ boxes: [b], xmin: b.xmin, xmax: b.xmax, bottom: b.ymax });
+    }
+  }
+  const columns: Point[][][] = [];
+  for (const col of cols) {
+    if (col.boxes.length >= 2) {
+      columns.push(col.boxes.map((b) => b.q)); // already top -> bottom
+    } else {
+      singles.push(col.boxes[0]!.q);
+    }
+  }
+  return { columns, singles };
+}
+
 /**
  * Resolves the content width (px) of a recognizer crop: the quad resized to the
  * recognizer height keeping aspect, clamped to the model's bucket width.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts
new file mode 100644
index 0000000000..4111db4057
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts
@@ -0,0 +1,555 @@
+// OCR pipeline engine: the per-page / per-box worklet functions and the
+// construction-time builders/validators behind `createOCR` (in ocr.ts). Split
+// out to keep the task file focused on the public API + factory wiring.
+// Everything here is internal — it is NOT re-exported from the package index.
+//
+// Worklet source-order rule: the worklet plugin captures referenced worklets in
+// source order, so a worklet must be defined BEFORE any worklet that calls it
+// (readStackedColumn -> detectQuads/recognizeGlyphStrip; readBoxVertical ->
+// readStackedColumn/recognizeQuad). The non-worklet builders/validators run at
+// construction time on the JS thread and have no such constraint.
+
+import { tensor, type Tensor } from '../../../core/tensor';
+import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
+import type { Model } from '../../../core/model';
+
+import type { ImageFormat } from '../image';
+import type { Point } from '../ops/points';
+import {
+  FORMAT_CONVERSION,
+  resize,
+  cvtColor,
+  toChannelsFirst,
+  normalize,
+  extractTextBoxes,
+  warpQuad,
+  ctcGreedyDecode,
+  type ColorConversionCode,
+  type CraftExtractOptions,
+  type DbnetExtractOptions,
+  type Quad,
+} from '../ops/image';
+import {
+  mapQuadToImage,
+  orderQuad,
+  quadSize,
+  contentWidthFor,
+  flattenQuad,
+  decodeGreedy,
+  ctcConfidence,
+  snapDetectBucket,
+  snapRecognizeBucket,
+  splitTallQuad,
+} from './ocrHelpers';
+
+// Detector input is raw RGB scaled to [0,1]; the per-architecture mean/std is
+// baked into the PTE, so the client only divides by 255.
+const DETECTOR_ALPHA = 1 / 255;
+const DETECTOR_BETA = 0;
+
+// Detector box-extraction tuning. These are stable across models and not worth
+// exposing as options — bake them in. (CRAFT = EasyOCR, DBNet = PaddleOCR.)
+const CRAFT_TEXT_THRESHOLD = 0.4;
+const CRAFT_LINK_THRESHOLD = 0.4;
+const CRAFT_LOW_TEXT_THRESHOLD = 0.7;
+const DBNET_BIN_THRESHOLD = 0.3;
+const DBNET_BOX_THRESHOLD = 0.6;
+const DBNET_UNCLIP_RATIO = 1.5;
+const DBNET_MIN_BOX_SIDE = 3;
+const DBNET_MAX_CANDIDATES = 1000;
+// DBNet exports a post-sigmoid prob map, so don't re-apply sigmoid (true only for
+// raw-logit heads).
+const APPLY_SIGMOID = false;
+
+// Stored at construction; CRAFT's `targetHeight` is per-run, so detectQuads adds it.
+export type DetectorExtractConfig = Omit<CraftExtractOptions, 'targetHeight'> | DbnetExtractOptions;
+
+// Per-detect-bucket scratch tensors, allocated once at construction (like RecSet)
+// and reused across the page pass and per-box re-detects. Only the source-resize
+// tensor depends on the run's input channel count, so detectQuads allocates that
+// one per call and reuses the rest from here.
+export type DetSet = {
+  readonly s: number;
+  readonly tColor: Tensor; // [s, s, 3]
+  readonly tCF: Tensor; // [3, s, s]
+  readonly tNorm: Tensor; // [3, s, s]
+  readonly tInput: Tensor; // [1, 3, s, s]
+  // The detector's output tensors (dbnet prob-map / craft heatmap+extras / a
+  // custom arch's raw outputs); the built-in decoder reads tOutputs[0].
+  readonly tOutputs: readonly Tensor[];
+};
+
+// Everything the detector pass needs, bundled so it can run both on the full
+// page and (for vertical text) on a single box crop to find its characters.
+export type DetectContext = {
+  readonly model: Model;
+  readonly detBuckets: readonly number[];
+  readonly numChannels: number;
+  readonly detCode: ColorConversionCode | null;
+  // Built-in decode config (craft/dbnet); undefined when a custom extractor is used.
+  readonly extractOpts?: DetectorExtractConfig;
+  // Custom TS box extractor for detectorKind 'custom'; takes precedence when set.
+  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
+  readonly detSets: ReadonlyMap<number, DetSet>;
+};
+
+// A recognizer width bucket's pre-allocated tensor-set (one per width in the
+// model's `recognize` buckets).
+export type RecSet = {
+  readonly width: number;
+  readonly tCanvas: Tensor;
+  readonly tCF: Tensor;
+  readonly tNorm: Tensor;
+  readonly tInput: Tensor;
+  readonly tLogits: Tensor;
+};
+
+// Recognizer state for reading one quad. The source image is passed per call,
+// so the same context reads both the full page and a per-box crop.
+export type RecContext = {
+  readonly model: Model;
+  readonly recSetByWidth: ReadonlyMap<number, RecSet>;
+  readonly recBuckets: readonly number[];
+  readonly recH: number;
+  readonly charset: string[];
+  // Per-model recognizer normalization / pad (resolved from OCROptions defaults).
+  readonly normAlpha: number | readonly number[];
+  readonly normBeta: number | readonly number[];
+  readonly padValue: number;
+  // Optional custom decode; falls back to greedy CTC when absent.
+  readonly decode?: (
+    logits: Tensor,
+    charset: readonly string[]
+  ) => { readonly text: string; readonly confidence: number };
+};
+
+// State the vertical-text path needs on top of `RecContext`: the detector (for
+// the second, character-level pass) and the page tensors it crops the box from.
+export type VerticalContext = {
+  readonly detCtx: DetectContext;
+  readonly rawPage: Tensor; // full page, native format — cropped per box for re-detection
+  readonly recCode: ColorConversionCode | null; // native format -> recognizer color
+  readonly recC: number;
+  readonly tallCropRatio: number; // height/width ratio above which a box is a column
+  // Per-page budget for the (expensive) stacked-column re-detection pass.
+  readonly redetectBudget: { remaining: number };
+};
+
+// cvtColor code to bring a source image format to RGB (both the detector and the
+// recognizer operate on RGB), or null when it already is RGB.
+export function toRgbCode(format: ImageFormat): ColorConversionCode | null {
+  'worklet';
+  return FORMAT_CONVERSION[format].rgb;
+}
+
+// Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
+// returns quads in `src` pixel space: letterbox into the snapped square bucket,
+// run `detect_<S>`, extract quads. Scratch is the bucket's cached DetSet; only
+// the source-resize tensor (input-channel-dependent) is allocated/freed here.
+export function detectQuads(
+  ctx: DetectContext,
+  src: Tensor,
+  width: number,
+  height: number,
+  charLevel = false
+): Point[][] {
+  'worklet';
+  const detS = snapDetectBucket(width, height, ctx.detBuckets);
+  // snapDetectBucket always returns one of detBuckets, so the set exists.
+  const ds = ctx.detSets.get(detS)!;
+  // Only the source resize depends on the run's channel count; the rest is cached.
+  const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
+  try {
+    src
+      .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
+      .throughIf(ctx.detCode !== null, cvtColor, ds.tColor, ctx.detCode!)
+      .through(toChannelsFirst, ds.tCF)
+      .through(normalize, ds.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
+      .copyTo(ds.tInput);
+
+    ctx.model.execute(`detect_${detS}`, [ds.tInput], [...ds.tOutputs]);
+    // A custom arch hands its raw outputs to the user extractor; the built-ins
+    // decode the heatmap (tOutputs[0]). CRAFT needs the per-run input height to
+    // restore its half-res boxes; `charLevel` switches to per-glyph boxes.
+    const quads = ctx.extractBoxes
+      ? ctx.extractBoxes(ds.tOutputs, detS)
+      : extractTextBoxes(
+          ds.tOutputs[0]!,
+          ctx.extractOpts!.mode === 'craft'
+            ? { ...ctx.extractOpts!, targetHeight: detS, charLevel }
+            : ctx.extractOpts!
+        );
+    return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
+  } finally {
+    tDetResize.dispose();
+  }
+}
+
+// Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap content width to a
+// recognizer bucket, warp -> normalize -> execute -> greedy-CTC decode.
+export function recognizeQuad(
+  ctx: RecContext,
+  src: Tensor,
+  corners: readonly Point[]
+): { text: string; conf: number } {
+  'worklet';
+  const cs = quadSize(corners);
+  const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
+  const desiredW = contentWidthFor(cs.width, cs.height, ctx.recH, maxRec);
+  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
+  // snapRecognizeBucket always returns one of recBuckets, so the set exists.
+  const rs = ctx.recSetByWidth.get(bucketW)!;
+  const contentWidth = Math.min(desiredW, bucketW);
+  warpQuad(src, rs.tCanvas, flattenQuad(corners), {
+    contentWidth,
+    align: 'left',
+    padMode: 'constant',
+    padValue: ctx.padValue,
+  });
+  rs.tCanvas
+    .through(toChannelsFirst, rs.tCF)
+    .through(normalize, rs.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
+    .copyTo(rs.tInput);
+  ctx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+  // A custom decode (e.g. attention/AR head) takes the raw logits; otherwise
+  // greedy CTC. Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
+  if (ctx.decode) {
+    const r = ctx.decode(rs.tLogits, ctx.charset);
+    return { text: r.text, conf: r.confidence };
+  }
+  const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
+  const text = decodeGreedy(indices, ctx.charset);
+  const conf = ctcConfidence(values, indices);
+  return { text, conf };
+}
+
+// Joins glyph quads (in `src` pixel space, in reading order) into one recognizer
+// strip — each glyph warped upright to the recognizer height and laid side by
+// side — and recognizes it as a single line (joint hconcat). Returns null when
+// nothing usable was assembled.
+//
+// Must be defined BEFORE its callers: the worklet plugin captures referenced
+// worklets in source order, so a forward reference is undefined at run time.
+export function recognizeGlyphStrip(
+  recCtx: RecContext,
+  src: Tensor,
+  recC: number,
+  glyphs: readonly (readonly Point[])[]
+): { text: string; conf: number } | null {
+  'worklet';
+  const recH = recCtx.recH;
+  const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+  // Warp each glyph upright to recognizer height (aspect preserved). A box that
+  // is much taller than wide is a merged run of stacked letters — split it into
+  // ~square single-letter cells first, so each lands in its own strip slot
+  // (otherwise N letters get squashed into one cell and read as garbage).
+  const slices: { tGlyph: Tensor; w: number }[] = [];
+  let totalW = 0;
+  for (const g of glyphs) {
+    const gsz = quadSize(g);
+    if (gsz.width < 1 || gsz.height < 1) {
+      continue;
+    }
+    const parts = Math.max(1, Math.round(gsz.height / Math.max(1, gsz.width)));
+    for (const cell of splitTallQuad(g, parts)) {
+      const gs = quadSize(cell);
+      if (gs.width < 1 || gs.height < 1) {
+        continue;
+      }
+      const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
+      const tGlyph = tensor('uint8', [recH, gw, recC]);
+      warpQuad(src, tGlyph, flattenQuad(cell), {
+        contentWidth: gw,
+        align: 'left',
+        padMode: 'constant',
+        padValue: recCtx.padValue,
+      });
+      slices.push({ tGlyph, w: gw });
+      totalW += gw;
+    }
+  }
+  if (slices.length === 0) {
+    return null;
+  }
+  try {
+    // Smallest bucket that fits the strip (snap up, no glyph truncated); widest
+    // bucket for very long columns.
+    const bucketW =
+      recCtx.recBuckets.find((w) => w >= totalW) ??
+      recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+    const rs = recCtx.recSetByWidth.get(bucketW)!;
+    // Assemble the strip row-major into the bucket canvas, neutral-padded.
+    const strip = new Uint8Array(recH * bucketW * recC);
+    strip.fill(recCtx.padValue);
+    let xOff = 0;
+    for (const s of slices) {
+      if (xOff >= bucketW) {
+        break;
+      }
+      const copyW = Math.min(s.w, bucketW - xOff);
+      const glyphBytes = new Uint8Array(recH * s.w * recC);
+      s.tGlyph.getData(glyphBytes);
+      for (let oy = 0; oy < recH; oy++) {
+        const srcStart = oy * s.w * recC;
+        const row = glyphBytes.subarray(srcStart, srcStart + copyW * recC);
+        strip.set(row, (oy * bucketW + xOff) * recC);
+      }
+      xOff += s.w;
+    }
+    rs.tCanvas.setData(strip);
+    rs.tCanvas
+      .through(toChannelsFirst, rs.tCF)
+      .through(normalize, rs.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
+      .copyTo(rs.tInput);
+    recCtx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+    if (recCtx.decode) {
+      const r = recCtx.decode(rs.tLogits, recCtx.charset);
+      return r.text.length > 0 ? { text: r.text, conf: r.confidence } : null;
+    }
+    const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
+    const text = decodeGreedy(indices, recCtx.charset);
+    const conf = ctcConfidence(values, indices);
+    return text.length > 0 ? { text, conf } : null;
+  } finally {
+    slices.forEach((s) => s.tGlyph.dispose());
+  }
+}
+
+// Reads a single tall box that packs several stacked glyphs the detector grouped
+// into one box (e.g. a tightly-set container code): crop it upright, re-detect
+// the glyphs (char-level pass), and read them top -> bottom as a joined strip.
+// Returns null — caller falls back to a horizontal read — when the box is tiny,
+// the per-page re-detect budget is spent, or nothing is found.
+export function readStackedColumn(
+  recCtx: RecContext,
+  vctx: VerticalContext,
+  ordered: readonly Point[],
+  size: { width: number; height: number }
+): { text: string; conf: number } | null {
+  'worklet';
+  const boxW = Math.round(size.width);
+  const boxH = Math.round(size.height);
+  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
+    return null;
+  }
+  vctx.redetectBudget.remaining--;
+  const recC = vctx.recC;
+  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
+  const tRecBox = tensor('uint8', [boxH, boxW, recC]);
+  try {
+    // Axis-aligned upright crop of the box, full width (no padding).
+    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
+      contentWidth: boxW,
+      align: 'left',
+      padMode: 'constant',
+      padValue: 0,
+    });
+    // Char-level second pass: per-glyph boxes (no grouping), in crop pixel space.
+    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
+    if (charQuads.length === 0) {
+      return null;
+    }
+    const boxSrc = vctx.recCode !== null ? cvtColor(tBoxRaw, tRecBox, vctx.recCode) : tBoxRaw;
+    // Stack reading order: top -> bottom by each glyph's upper edge.
+    const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
+    return recognizeGlyphStrip(recCtx, boxSrc, recC, glyphs);
+  } finally {
+    tBoxRaw.dispose();
+    tRecBox.dispose();
+  }
+}
+
+// Reads one box that may be an upright stacked column (e.g. a shipping-container
+// code, letters stacked top-to-bottom). A normal-aspect box is read horizontally;
+// a tall box is read as a stacked column, falling back to the horizontal read.
+// `stacked` reports whether the column path produced the read (caller applies the
+// lenient vertical drop-score to those).
+export function readBoxVertical(
+  recCtx: RecContext,
+  vctx: VerticalContext,
+  pageSrc: Tensor,
+  ordered: readonly Point[],
+  size: { width: number; height: number }
+): { text: string; conf: number; stacked: boolean } {
+  'worklet';
+  if (size.height >= size.width * vctx.tallCropRatio) {
+    const stacked = readStackedColumn(recCtx, vctx, ordered, size);
+    if (stacked) {
+      return { ...stacked, stacked: true };
+    }
+  }
+  return { ...recognizeQuad(recCtx, pageSrc, ordered), stacked: false };
+}
+
+// Validates the detector methods against the architecture's input/output spec
+// and returns CRAFT's extra-output channel counts (empty for dbnet/custom). For
+// 'custom' only the shared RGB input contract is enforced — outputs are read from
+// the PTE metadata and handed to the user's extractBoxes. Throws on any mismatch.
+export function validateDetectorSchema(
+  model: Model,
+  detBuckets: readonly number[],
+  detectorKind: 'craft' | 'dbnet' | 'custom',
+  extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[]
+): number[] {
+  // CRAFT's half-resolution heatmap needs even detect-bucket sides.
+  if (detectorKind === 'craft' && detBuckets.some((s) => s % 2 !== 0)) {
+    throw new Error('OCR: every CRAFT detect bucket side must be even (half-resolution heatmap).');
+  }
+  const detInSpec = [SymbolicTensor('float32', [1, 3, 'H', 'W'])];
+  if (detectorKind === 'custom') {
+    if (!extractBoxes) {
+      throw new Error("OCR: detectorKind 'custom' requires an extractBoxes worklet.");
+    }
+    // Outputs are arbitrary (read from metadata, handed to extractBoxes); only
+    // the shared RGB input contract is enforced. getMethodMeta throws if missing.
+    for (const s of detBuckets) {
+      const inShape = model.getMethodMeta(`detect_${s}`).inputTensorMeta[0]?.shape;
+      if (!inShape || inShape.length !== 4 || inShape[1] !== 3) {
+        throw new Error(`OCR: detect_${s} must take a [1, 3, ${s}, ${s}] RGB input.`);
+      }
+    }
+    return [];
+  }
+  // Validate every detect bucket against the architecture's output spec; keep
+  // the largest bucket's meta for the constant CRAFT extra-output channels.
+  const detOutSpec =
+    detectorKind === 'dbnet'
+      ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
+      : [
+          SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
+          SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
+        ];
+  const detMeta = validateModelSchema(
+    model,
+    `detect_${detBuckets[detBuckets.length - 1]}`,
+    detInSpec,
+    detOutSpec
+  );
+  for (let i = 0; i < detBuckets.length - 1; i++) {
+    validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
+  }
+  // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
+  return detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
+}
+
+// Built-in box-decode config for a detector kind (custom archs decode in TS via
+// extractBoxes instead, so they get `undefined`).
+export function buildExtractOpts(
+  detectorKind: 'craft' | 'dbnet' | 'custom'
+): DetectorExtractConfig | undefined {
+  if (detectorKind === 'custom') {
+    return undefined;
+  }
+  return detectorKind === 'dbnet'
+    ? {
+        mode: 'dbnet',
+        binThreshold: DBNET_BIN_THRESHOLD,
+        boxThreshold: DBNET_BOX_THRESHOLD,
+        unclipRatio: DBNET_UNCLIP_RATIO,
+        minBoxSide: DBNET_MIN_BOX_SIDE,
+        maxCandidates: DBNET_MAX_CANDIDATES,
+        applySigmoid: APPLY_SIGMOID,
+      }
+    : {
+        mode: 'craft',
+        textThreshold: CRAFT_TEXT_THRESHOLD,
+        linkThreshold: CRAFT_LINK_THRESHOLD,
+        lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
+      };
+}
+
+// Pre-allocates one detector scratch-set per detect bucket (channel-independent
+// tensors; the per-run source-resize tensor is allocated in detectQuads). Mirrors
+// buildRecognizerSets — runs at construction.
+export function buildDetectorSets(
+  model: Model,
+  detBuckets: readonly number[],
+  detectorKind: 'craft' | 'dbnet' | 'custom',
+  detExtraChannels: readonly number[]
+): DetSet[] {
+  return detBuckets.map((s) => {
+    const heat = s / 2;
+    // Custom archs declare arbitrary outputs — size them straight from the PTE's
+    // method metadata. Built-ins keep their known heatmap (+ craft extras) shapes.
+    let tOutputs: Tensor[];
+    if (detectorKind === 'custom') {
+      tOutputs = model
+        .getMethodMeta(`detect_${s}`)
+        .outputTensorMeta.map((m) => tensor(m.dtype, m.shape));
+    } else {
+      const tHeatmap =
+        detectorKind === 'dbnet'
+          ? tensor('float32', [1, 1, s, s])
+          : tensor('float32', [1, heat, heat, 2]);
+      tOutputs = [tHeatmap, ...detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]))];
+    }
+    return {
+      s,
+      tColor: tensor('uint8', [s, s, 3]),
+      tCF: tensor('uint8', [3, s, s]),
+      tNorm: tensor('float32', [3, s, s]),
+      tInput: tensor('float32', [1, 3, s, s]),
+      tOutputs,
+    };
+  });
+}
+
+// Pre-allocates one recognizer tensor-set per width bucket (each `recognize_<W>`
+// validated once) and derives the constant channel/height/vocab contract from the
+// first bucket. Kept out of the task factory; runs at construction.
+export function buildRecognizerSets(
+  model: Model,
+  recBuckets: readonly number[]
+): { sets: RecSet[]; recC: number; recH: number; vocabSize: number } {
+  let recC = 0;
+  let recH = 0;
+  let vocabSize = 0;
+  const sets = recBuckets.map((w, i) => {
+    const m = validateModelSchema(
+      model,
+      `recognize_${w}`,
+      [SymbolicTensor('float32', [1, 'C', 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 'T', 'V'])]
+    );
+    const inShape = m.inputTensorMeta[0]!.shape;
+    if (i === 0) {
+      // Channels/height/vocab are constant across the width buckets.
+      recC = inShape[1]!;
+      recH = inShape[2]!;
+      vocabSize = m.outputTensorMeta[0]!.shape[2]!;
+    }
+    return {
+      width: w,
+      tCanvas: tensor('uint8', [recH, w, recC]),
+      tCF: tensor('uint8', [recC, recH, w]),
+      tNorm: tensor('float32', [recC, recH, w]),
+      tInput: tensor('float32', inShape),
+      tLogits: tensor('float32', m.outputTensorMeta[0]!.shape),
+    };
+  });
+  return { sets, recC, recH, vocabSize };
+}
+
+// Frees a detector scratch-set's tensors (input prep + per-bucket outputs).
+export function disposeDetSets(detSets: readonly DetSet[]): void {
+  for (const d of detSets) {
+    d.tColor.dispose();
+    d.tCF.dispose();
+    d.tNorm.dispose();
+    d.tInput.dispose();
+    for (const t of d.tOutputs) {
+      t.dispose();
+    }
+  }
+}
+
+// Frees a recognizer scratch-set's tensors.
+export function disposeRecSets(recSets: readonly RecSet[]): void {
+  for (const s of recSets) {
+    s.tCanvas.dispose();
+    s.tCF.dispose();
+    s.tNorm.dispose();
+    s.tInput.dispose();
+    s.tLogits.dispose();
+  }
+}

From 557798c02dff322e91b2f3847d14e2a93432adb5 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 17:47:10 +0200
Subject: [PATCH 18/29] [RNE Rewrite] refactor(cv): share JSI option-readers,
 move warpQuad to image_ops
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

ocr_ops.cpp held two things that aren't OCR-specific:

- The JSI option-readers (getNumberProp/getStringProp/getBoolProp/getBoolPropOr)
  are generic plumbing — promote them to utils.h so image_ops/ocr_ops share one
  copy instead of re-rolling the same hasProperty/isX pattern.
- warpQuad is a generic perspective-crop image op (getPerspectiveTransform +
  warpPerspective + pad/align), no OCR in the math. Move it to image_ops.cpp next
  to resize/cvtColor/gridSample; update headers + install.cpp wiring.

ocr_ops.cpp 835 -> 679 lines; it now holds only OCR detector/sequence
postprocessing (CRAFT grouping, DBNet contouring, CTC argmax). Verified with
check-cpp-warnings.sh (clang++ -fsyntax-only vs ExecuTorch/JSI/OpenCV): clean.
---
 .../cpp/extensions/cv/image_ops.cpp           | 129 ++++++++++++++
 .../cpp/extensions/cv/image_ops.h             |   4 +
 .../cpp/extensions/cv/install.cpp             |   2 +-
 .../cpp/extensions/cv/ocr_ops.cpp             | 158 +-----------------
 .../cpp/extensions/cv/ocr_ops.h               |   4 -
 .../cpp/extensions/cv/utils.h                 |  41 +++++
 6 files changed, 176 insertions(+), 162 deletions(-)

diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index d9b0d26404..761553106b 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -1,6 +1,7 @@
 #include "image_ops.h"
 
 #include <algorithm>
+#include <array>
 #include <cmath>
 #include <numeric>
 #include <stdexcept>
@@ -819,4 +820,132 @@ void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
     };
     module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
 }
+
+// ------------------------------- warpQuad ----------------------------------
+// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
+// resize-to-height + pad/align). A generic image op; used by the OCR recognizer.
+void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
+    auto name = "warpQuad";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 4) {
+            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
+        }
+        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
+        }
+        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
+            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
+        }
+        if (!args[3].isObject()) {
+            throw jsi::JSError(rt, "warpQuad: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
+        }
+        auto quadArr = args[2].asObject(rt).asArray(rt);
+        auto opts = args[3].asObject(rt);
+
+        if (quadArr.length(rt) != 8) {
+            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
+        }
+        if (src->shape_[2] != dst->shape_[2]) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
+        }
+
+        const int32_t channels = src->shape_[2];
+        const int32_t recH = dst->shape_[0];
+        const int32_t bucketW = dst->shape_[1];
+
+        if (!opts.hasProperty(rt, "contentWidth") ||
+            !opts.getProperty(rt, "contentWidth").isNumber()) {
+            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
+        }
+        const int32_t contentWidth =
+            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
+                       bucketW);
+        const std::string padMode = getStringProp(rt, opts, "padMode");
+        const double padValue = getNumberProp(rt, opts, "padValue");
+        const std::string align = getStringProp(rt, opts, "align");
+
+        std::array<::cv::Point2f, 4> quad;
+        for (std::size_t i = 0; i < 8; ++i) {
+            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
+                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
+            }
+        }
+        for (std::size_t i = 0; i < 4; ++i) {
+            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
+                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
+        }
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
+        }
+        if (!src->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
+        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
+
+        try {
+            const ::cv::Point2f dstPts[4] = {{0.0f, 0.0f},
+                                             {static_cast<float>(contentWidth), 0.0f},
+                                             {static_cast<float>(contentWidth),
+                                              static_cast<float>(recH)},
+                                             {0.0f, static_cast<float>(recH)}};
+            const ::cv::Point2f srcPts[4] = {quad[0], quad[1], quad[2], quad[3]};
+            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts, dstPts);
+            ::cv::Mat content;
+            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
+                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
+
+            ::cv::Scalar padColor;
+            if (padMode == "cornerMean") {
+                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
+                ::cv::Scalar acc(0, 0, 0, 0);
+                const std::array<::cv::Rect, 4> rects = {
+                    ::cv::Rect(0, 0, patch, patch),
+                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
+                    ::cv::Rect(0, recH - patch, patch, patch),
+                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
+                for (const auto &r : rects) {
+                    acc += ::cv::mean(content(r));
+                }
+                padColor = acc / 4.0;
+            } else {
+                padColor = ::cv::Scalar::all(padValue);
+            }
+
+            dstMat.setTo(padColor);
+            const int32_t offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
+            content.copyTo(dstMat(::cv::Rect(offsetX, 0, contentWidth, recH)));
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[1]);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             4, fnBody));
+}
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
index dc24873a66..90542dd212 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
@@ -10,4 +10,8 @@ void install_toChannelsLast(facebook::jsi::Runtime &rt, facebook::jsi::Object &m
 void install_normalize(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_applyColormap(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_gridSample(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
+// resize-to-height + pad/align), HWC uint8. Generic image op; the OCR recognizer
+// uses it to lift each text quad into a fixed-height recognizer canvas.
+void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/install.cpp b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
index 9fd35bdc22..7377f0d3ce 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/install.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
@@ -16,12 +16,12 @@ void install(facebook::jsi::Runtime &rt, facebook::jsi::Object &module) {
     image_ops::install_normalize(rt, cvModule);
     image_ops::install_applyColormap(rt, cvModule);
     image_ops::install_gridSample(rt, cvModule);
+    image_ops::install_warpQuad(rt, cvModule);
 
     box_ops::install_nms(rt, cvModule);
     box_ops::install_restrictToBox(rt, cvModule);
 
     ocr_ops::install_extractTextBoxes(rt, cvModule);
-    ocr_ops::install_warpQuad(rt, cvModule);
     ocr_ops::install_ctcGreedyDecode(rt, cvModule);
 
     module.setProperty(rt, "cv", cvModule);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 10f3765af1..893e20a7b1 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -20,6 +20,7 @@
 
 #include "core/dtype.h"
 #include "core/tensor.h"
+#include "utils.h"
 
 // Detector postprocessing geometry: CRAFT text-map grouping + DBNet prob-map ->
 // oriented quads. Pure OpenCV, kept native. ctcGreedyDecode (per-timestep argmax
@@ -509,37 +510,6 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, floa
     return quads;
 }
 
-// ----------------------------- option readers ------------------------------
-// Required option getters (defaults live in the TS wrappers, so these throw).
-double getNumberProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isNumber()) {
-        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a number");
-    }
-    return opts.getProperty(rt, name).asNumber();
-}
-
-std::string getStringProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isString()) {
-        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a string");
-    }
-    return opts.getProperty(rt, name).asString(rt).utf8(rt);
-}
-
-bool getBoolProp(jsi::Runtime &rt, const jsi::Object &opts, const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
-        throw jsi::JSError(rt, std::string("options.") + name + " is required and must be a boolean");
-    }
-    return opts.getProperty(rt, name).asBool();
-}
-
-// Optional boolean (defaults when absent) — used for flags a caller may omit.
-bool getBoolPropOr(jsi::Runtime &rt, const jsi::Object &opts, const char *name, bool fallback) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
-        return fallback;
-    }
-    return opts.getProperty(rt, name).asBool();
-}
-
 // Flatten quads to a JS double array, 10 per box (x0,y0..x3,y3,score,angle).
 jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
     jsi::Array out(rt, quads.size() * 10);
@@ -638,132 +608,6 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                                                              2, fnBody));
 }
 
-// ------------------------------- warpQuad ----------------------------------
-void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
-    auto name = "warpQuad";
-    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
-                     size_t count) -> jsi::Value {
-        if (count != 4) {
-            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
-        }
-        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
-        }
-        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
-        }
-        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
-            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
-        }
-        if (!args[3].isObject()) {
-            throw jsi::JSError(rt, "warpQuad: options must be an object");
-        }
-        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
-        if (src.get() == dst.get()) {
-            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
-        }
-        auto quadArr = args[2].asObject(rt).asArray(rt);
-        auto opts = args[3].asObject(rt);
-
-        if (quadArr.length(rt) != 8) {
-            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
-        }
-        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
-        }
-        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
-            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
-        }
-        if (src->shape_[2] != dst->shape_[2]) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
-        }
-
-        const int32_t channels = src->shape_[2];
-        const int32_t recH = dst->shape_[0];
-        const int32_t bucketW = dst->shape_[1];
-
-        if (!opts.hasProperty(rt, "contentWidth") ||
-            !opts.getProperty(rt, "contentWidth").isNumber()) {
-            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
-        }
-        const int32_t contentWidth =
-            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
-                       bucketW);
-        const std::string padMode = getStringProp(rt, opts, "padMode");
-        const double padValue = getNumberProp(rt, opts, "padValue");
-        const std::string align = getStringProp(rt, opts, "align");
-
-        std::array<::cv::Point2f, 4> quad;
-        for (std::size_t i = 0; i < 8; ++i) {
-            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
-                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
-            }
-        }
-        for (std::size_t i = 0; i < 4; ++i) {
-            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
-                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
-        }
-
-        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
-        if (!srcLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
-        }
-        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
-        if (!dstLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
-        }
-        if (!src->data_ || !dst->data_) {
-            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
-        }
-
-        const int cvType = CV_MAKETYPE(CV_8U, channels);
-        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
-        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
-
-        try {
-            const ::cv::Point2f dstPts[4] = {{0.0f, 0.0f},
-                                             {static_cast<float>(contentWidth), 0.0f},
-                                             {static_cast<float>(contentWidth),
-                                              static_cast<float>(recH)},
-                                             {0.0f, static_cast<float>(recH)}};
-            const ::cv::Point2f srcPts[4] = {quad[0], quad[1], quad[2], quad[3]};
-            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts, dstPts);
-            ::cv::Mat content;
-            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
-                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
-
-            ::cv::Scalar padColor;
-            if (padMode == "cornerMean") {
-                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
-                ::cv::Scalar acc(0, 0, 0, 0);
-                const std::array<::cv::Rect, 4> rects = {
-                    ::cv::Rect(0, 0, patch, patch),
-                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
-                    ::cv::Rect(0, recH - patch, patch, patch),
-                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
-                for (const auto &r : rects) {
-                    acc += ::cv::mean(content(r));
-                }
-                padColor = acc / 4.0;
-            } else {
-                padColor = ::cv::Scalar::all(padValue);
-            }
-
-            dstMat.setTo(padColor);
-            const int32_t offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
-            content.copyTo(dstMat(::cv::Rect(offsetX, 0, contentWidth, recH)));
-        } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
-        }
-        return jsi::Value(rt, args[1]);
-    };
-    module.setProperty(rt, name,
-                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
-                                                             4, fnBody));
-}
-
 // --------------------------- ctcGreedyDecode -------------------------------
 void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
     auto name = "ctcGreedyDecode";
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
index dc7905fe99..893123fec3 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
@@ -7,10 +7,6 @@ namespace rnexecutorch::extensions::cv::ocr_ops {
 // x0,y0..x3,y3,score,angle), in detector-input pixels.
 void install_extractTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 
-// Perspective-crop an oriented quad of `src` into the recognizer canvas `dst`
-// (crop + resize-to-height + bucket-pad), HWC uint8.
-void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-
 // Per-timestep argmax + max value over [..,T,V] logits -> flat [idx,prob,...];
 // options.softmax makes `prob` a probability.
 void install_ctcGreedyDecode(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/utils.h b/packages/react-native-executorch/cpp/extensions/cv/utils.h
index a29961d0b6..75a9de194a 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/utils.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/utils.h
@@ -1,8 +1,10 @@
 #pragma once
 
 #include "core/dtype.h"
+#include <jsi/jsi.h>
 #include <opencv2/core.hpp>
 #include <stdexcept>
+#include <string>
 
 namespace rnexecutorch::extensions::cv {
 
@@ -18,4 +20,43 @@ inline int dtypeToCvDepth(rnexecutorch::core::types::DType dtype) {
     throw std::invalid_argument("unsupported dtype");
 }
 
+// ------------------------------ JSI option readers ------------------------------
+// Required option getters shared by the cv op installers. Defaults live in the TS
+// wrappers, so a missing/mis-typed property is a contract violation -> throw.
+inline double getNumberProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
+                            const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isNumber()) {
+        throw facebook::jsi::JSError(rt, std::string("options.") + name +
+                                             " is required and must be a number");
+    }
+    return opts.getProperty(rt, name).asNumber();
+}
+
+inline std::string getStringProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
+                                 const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isString()) {
+        throw facebook::jsi::JSError(rt, std::string("options.") + name +
+                                             " is required and must be a string");
+    }
+    return opts.getProperty(rt, name).asString(rt).utf8(rt);
+}
+
+inline bool getBoolProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
+                        const char *name) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
+        throw facebook::jsi::JSError(rt, std::string("options.") + name +
+                                             " is required and must be a boolean");
+    }
+    return opts.getProperty(rt, name).asBool();
+}
+
+// Optional boolean (defaults when absent) — used for flags a caller may omit.
+inline bool getBoolPropOr(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
+                          const char *name, bool fallback) {
+    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
+        return fallback;
+    }
+    return opts.getProperty(rt, name).asBool();
+}
+
 } // namespace rnexecutorch::extensions::cv

From be447a1488d8a04c4cfc6e83e4751e6b62bd040f Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Tue, 30 Jun 2026 17:47:25 +0200
Subject: [PATCH 19/29] [RNE Rewrite] docs(ocr): fix stale models.ts comment;
 revert VIEW_HEIGHT rename
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- models.ts: the OCR family comment claimed the recognizer profile (norm, color,
  padding, CTC, confidence) is 'derived from detectorKind' — stale. detectorKind
  only selects the box decoder + default drop score; the rest is the shared baked
  contract, now overridable via recognizerNorm/recognizerPadValue/decode.
- ImageViewport: keep VIEW_HEIGHT as-is (drop the DEFAULT_VIEW_HEIGHT rename) — the
  per-instance override already flows through viewHeight, so the rename was churn.
---
 apps/computer-vision/components/ImageViewport.tsx |  4 ++--
 packages/react-native-executorch/src/models.ts    | 11 +++++++----
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/apps/computer-vision/components/ImageViewport.tsx b/apps/computer-vision/components/ImageViewport.tsx
index e92684f43c..cd258f1528 100644
--- a/apps/computer-vision/components/ImageViewport.tsx
+++ b/apps/computer-vision/components/ImageViewport.tsx
@@ -12,7 +12,7 @@ import {
 import { theme } from '../theme';
 
 const VIEW_WIDTH = Dimensions.get('window').width - 32;
-const DEFAULT_VIEW_HEIGHT = Math.round((VIEW_WIDTH * 16) / 9);
+const VIEW_HEIGHT = Math.round((VIEW_WIDTH * 16) / 9);
 
 /** A 2D point in the displayed image's pixel coordinates. */
 type Point = { readonly x: number; readonly y: number };
@@ -44,7 +44,7 @@ export function ImageViewport({
   height,
   boxes,
 }: ImageViewportProps) {
-  const viewHeight = height ?? DEFAULT_VIEW_HEIGHT;
+  const viewHeight = height ?? VIEW_HEIGHT;
 
   // Map image-pixel polygons into canvas space using the same contain-fit
   // transform Skia uses to draw the image, then build one stroked path.
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 1c3e5c1ce2..93dd03689c 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -541,10 +541,13 @@ const YOLO26_XLARGE_SEG_640_XNNPACK_FP32: InstanceSegmenterModel<'xyxy', CocoCla
 const ALL_MINILM_L6_V2_TOKENIZER = `${BASE_URL}-all-MiniLM-L6-v2/${VERSION_TAG}/tokenizer.json`;
 // OCR
 // =============================================================================
-// EasyOCR (CRAFT + CRNN) and PaddleOCR (DBNet + SVTR). The detector/recognizer
-// profile (normalization, color, padding, CTC head, confidence) is derived from
-// `detectorKind` inside the task, so each model only declares its architecture,
-// input-size `buckets`, and `charset`. `charset` for EasyOCR is set per language.
+// EasyOCR (CRAFT + CRNN) and PaddleOCR (DBNet + SVTR). `detectorKind` selects the
+// box decoder (CRAFT heatmap grouping vs DBNet prob-map contouring) and the default
+// drop score; everything else — RGB input, recognizer normalization/padding, CTC
+// decode, confidence — is the shared baked contract (overridable per model via
+// `recognizerNorm`/`recognizerPadValue`/`decode`, which these built-ins leave at
+// the defaults). So each only declares its architecture, input-size `buckets`, and
+// `charset`. `charset` for EasyOCR is set per language.
 const EASYOCR_OPTS: OCROptions = {
   detectorKind: 'craft', // CRAFT: text + affinity heatmaps grouped into lines
   charset: alphabets.english, // overridden per language

From 60801206ef52e49822d2f0532f99e2039f681ba6 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Wed, 1 Jul 2026 09:18:10 +0200
Subject: [PATCH 20/29] [RNE Rewrite] satisfy clang-tidy on ocr_ops/image_ops

---
 .../cpp/extensions/cv/image_ops.cpp           | 24 +++---
 .../cpp/extensions/cv/ocr_ops.cpp             | 77 +++++++++++--------
 2 files changed, 55 insertions(+), 46 deletions(-)

diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index 761553106b..31cef5025c 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -717,7 +717,7 @@ void install_applyColormap(jsi::Runtime &rt, jsi::Object &module) {
 }
 
 void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
-    auto name = "gridSample";
+    const auto *name = "gridSample";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
         if (count != 3) {
             throw jsi::JSError(rt, "Usage: gridSample(src, grid, dst)");
@@ -771,7 +771,7 @@ void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
         const int32_t gridH = gs[gs.size() - 2];
         const int32_t gridW = gs[gs.size() - 1];
         const int32_t plane = gridH * gridW;
-        const float *g = reinterpret_cast<const float *>(grid->data_.get());
+        const auto *g = reinterpret_cast<const float *>(grid->data_.get());
 
         // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
         auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
@@ -795,8 +795,8 @@ void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
             const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
                                          static_cast<float>(gridH - 1)
                                    : 0.0f;
-            float *rowX = mapX.ptr<float>(oy);
-            float *rowY = mapY.ptr<float>(oy);
+            auto *rowX = mapX.ptr<float>(oy);
+            auto *rowY = mapY.ptr<float>(oy);
             for (int32_t ox = 0; ox < w; ++ox) {
                 const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
                                              static_cast<float>(gridW - 1)
@@ -825,7 +825,7 @@ void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
 // Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
 // resize-to-height + pad/align). A generic image op; used by the OCR recognizer.
 void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
-    auto name = "warpQuad";
+    const auto *name = "warpQuad";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
                      size_t count) -> jsi::Value {
         if (count != 4) {
@@ -908,13 +908,13 @@ void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
         ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
 
         try {
-            const ::cv::Point2f dstPts[4] = {{0.0f, 0.0f},
-                                             {static_cast<float>(contentWidth), 0.0f},
-                                             {static_cast<float>(contentWidth),
-                                              static_cast<float>(recH)},
-                                             {0.0f, static_cast<float>(recH)}};
-            const ::cv::Point2f srcPts[4] = {quad[0], quad[1], quad[2], quad[3]};
-            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts, dstPts);
+            const std::array<::cv::Point2f, 4> dstPts = {
+                ::cv::Point2f{0.0f, 0.0f},
+                {static_cast<float>(contentWidth), 0.0f},
+                {static_cast<float>(contentWidth), static_cast<float>(recH)},
+                {0.0f, static_cast<float>(recH)}};
+            const std::array<::cv::Point2f, 4> srcPts = {quad[0], quad[1], quad[2], quad[3]};
+            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts.data(), dstPts.data());
             ::cv::Mat content;
             ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
                                   ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 893e20a7b1..77401e6c23 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -6,6 +6,7 @@
 #include <cstddef>
 #include <jsi/jsi.h>
 #include <limits>
+#include <numbers>
 #include <numeric>
 #include <opencv2/core/check.hpp>
 #include <optional>
@@ -33,10 +34,10 @@ using TensorHostObject = rnexecutorch::core::tensor::TensorHostObject;
 namespace {
 // ----------------------------- geometry types ------------------------------
 struct Box {
-    float x0, y0, x1, y1; // axis-aligned (p1=min, p2=max)
+    float x0{}, y0{}, x1{}, y1{}; // axis-aligned (p1=min, p2=max)
     float angle = 0.0f;
-    float width() const { return x1 - x0; }
-    float height() const { return y1 - y0; }
+    [[nodiscard]] float width() const { return x1 - x0; }
+    [[nodiscard]] float height() const { return y1 - y0; }
 };
 
 struct Quad {
@@ -72,8 +73,8 @@ std::pair<::cv::Mat, ::cv::Mat> interleavedToMats(std::span<const float> data, :
     ::cv::Mat affinityMap(size, CV_32F);
     const auto w = static_cast<std::size_t>(size.width);
     for (std::size_t i = 0; i < data.size(); ++i) {
-        const int32_t x = static_cast<int32_t>((i / 2) % w);
-        const int32_t y = static_cast<int32_t>((i / 2) / w);
+        const auto x = static_cast<int32_t>((i / 2) % w);
+        const auto y = static_cast<int32_t>((i / 2) / w);
         if (i % 2 == 0) {
             textMap.at<float>(y, x) = data[i];
         } else {
@@ -89,7 +90,7 @@ void dilateComponent(::cv::Mat &segMap, const ::cv::Mat &stats, int32_t i, int32
     const int32_t y = stats.at<int32_t>(i, ::cv::CC_STAT_TOP);
     const int32_t w = stats.at<int32_t>(i, ::cv::CC_STAT_WIDTH);
     const int32_t h = stats.at<int32_t>(i, ::cv::CC_STAT_HEIGHT);
-    const int32_t dilationRadius =
+    const auto dilationRadius =
         static_cast<int32_t>(std::sqrt(static_cast<double>(area) / std::max(w, h)) * 2);
     const int32_t sx = std::max(x - dilationRadius, 0);
     const int32_t ex = std::min(x + w + dilationRadius, imgW);
@@ -109,7 +110,7 @@ std::optional<Box> boxFromComponent(const ::cv::Mat &textMap, const ::cv::Mat &l
         return std::nullopt;
     }
     ::cv::Mat mask = (labels == i);
-    double maxVal;
+    double maxVal = 0.0;
     ::cv::minMaxLoc(textMap, nullptr, &maxVal, nullptr, nullptr, mask);
     if (maxVal < static_cast<double>(lowTextThreshold)) {
         return std::nullopt;
@@ -140,7 +141,8 @@ std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityM
                                         float lowTextThreshold) {
     const int32_t imgH = textMap.rows;
     const int32_t imgW = textMap.cols;
-    ::cv::Mat textScore, affinityScore;
+    ::cv::Mat textScore;
+    ::cv::Mat affinityScore;
     ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
     ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
                     ::cv::THRESH_BINARY);
@@ -149,7 +151,9 @@ std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityM
     ::cv::Mat binary;
     comb.convertTo(binary, CV_8UC1);
 
-    ::cv::Mat labels, stats, centroids;
+    ::cv::Mat labels;
+    ::cv::Mat stats;
+    ::cv::Mat centroids;
     const int32_t nLabels = ::cv::connectedComponentsWithStats(binary, labels, stats, centroids, 4);
 
     std::vector<Box> boxes;
@@ -174,7 +178,7 @@ std::tuple<float, float, bool> fitLineToShortestSides(const Box &b, float vertic
         sides[static_cast<std::size_t>(i)] = {dist(p1, p2), i};
         mids[static_cast<std::size_t>(i)] = {(p1.x + p2.x) * 0.5f, (p1.y + p2.y) * 0.5f};
     }
-    std::sort(sides.begin(), sides.end());
+    std::ranges::sort(sides);
     ::cv::Point2f m1 = mids[static_cast<std::size_t>(sides[0].second)];
     ::cv::Point2f m2 = mids[static_cast<std::size_t>(sides[1].second)];
     const bool isVertical = std::fabs(m2.x - m1.x) < verticalThreshold;
@@ -193,7 +197,7 @@ std::tuple<float, float, bool> fitLineToShortestSides(const Box &b, float vertic
 
 Box rotateBox(const Box &b, float angleDeg) {
     const ::cv::Point2f ctr = center(b);
-    const float rad = angleDeg * static_cast<float>(M_PI) / 180.0f;
+    const float rad = angleDeg * std::numbers::pi_v<float> / 180.0f;
     float minX = std::numeric_limits<float>::max();
     float minY = std::numeric_limits<float>::max();
     float maxX = std::numeric_limits<float>::lowest();
@@ -205,7 +209,7 @@ Box rotateBox(const Box &b, float angleDeg) {
         maxX = std::max(maxX, r.x);
         maxY = std::max(maxY, r.y);
     }
-    return {minX, minY, maxX, maxY, b.angle};
+    return {.x0 = minX, .y0 = minY, .x1 = maxX, .y1 = maxY, .angle = b.angle};
 }
 
 float minDistanceBetween(const Box &a, const Box &b) {
@@ -250,8 +254,7 @@ findClosestBox(const std::vector<Box> &boxes, const std::unordered_set<std::size
 }
 
 Box mergeBoxes(const Box &a, const Box &b) {
-    return {std::min(a.x0, b.x0), std::min(a.y0, b.y0), std::max(a.x1, b.x1), std::max(a.y1, b.y1),
-            a.angle};
+    return {.x0 = std::min(a.x0, b.x0), .y0 = std::min(a.y0, b.y0), .x1 = std::max(a.x1, b.x1), .y1 = std::max(a.y1, b.y1), .angle = a.angle};
 }
 
 // CRAFT box grouping -> reading-ordered text lines.
@@ -259,8 +262,8 @@ std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
                                 float distanceThreshold, float heightThreshold,
                                 float minSideThreshold, float maxSideThreshold,
                                 float verticalThreshold) {
-    std::sort(boxes.begin(), boxes.end(),
-              [](const Box &a, const Box &b) { return maxSide(a) > maxSide(b); });
+    std::ranges::sort(boxes,
+                      [](const Box &a, const Box &b) { return maxSide(a) > maxSide(b); });
 
     std::vector<Box> merged;
     std::unordered_set<std::size_t> ignored;
@@ -273,7 +276,7 @@ std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
 
         while (true) {
             auto [slope, intercept, isVertical] = fitLineToShortestSides(current, verticalThreshold);
-            lineAngle = isVertical ? -90.0f : std::atan(slope) * 180.0f / static_cast<float>(M_PI);
+            lineAngle = isVertical ? -90.0f : std::atan(slope) * 180.0f / std::numbers::pi_v<float>;
             auto closest =
                 findClosestBox(boxes, ignored, current, isVertical, slope, intercept, centerThreshold);
             if (!closest) {
@@ -309,8 +312,8 @@ std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
     }
 
     // reading order: rows by top-Y, then left-to-right within a row
-    std::sort(filtered.begin(), filtered.end(),
-              [](const Box &a, const Box &b) { return a.y0 < b.y0; });
+    std::ranges::sort(filtered,
+                      [](const Box &a, const Box &b) { return a.y0 < b.y0; });
     float yThresh = 0.0f;
     if (!filtered.empty()) {
         float total = 0.0f;
@@ -340,7 +343,8 @@ std::vector<Box> getCharBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinity
                                          float lowTextThreshold) {
     const int32_t imgH = textMap.rows;
     const int32_t imgW = textMap.cols;
-    ::cv::Mat textScore, affinityScore;
+    ::cv::Mat textScore;
+    ::cv::Mat affinityScore;
     ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
     ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
                     ::cv::THRESH_BINARY);
@@ -353,7 +357,9 @@ std::vector<Box> getCharBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinity
 
     ::cv::Mat binary;
     comb.convertTo(binary, CV_8UC1);
-    ::cv::Mat labels, stats, centroids;
+    ::cv::Mat labels;
+    ::cv::Mat stats;
+    ::cv::Mat centroids;
     const int32_t nLabels = ::cv::connectedComponentsWithStats(binary, labels, stats, centroids, 4);
 
     std::vector<Box> boxes;
@@ -408,7 +414,7 @@ std::vector<Quad> extractCraft(std::span<const float> data, int32_t heatW, int32
         // can take it.
         const ::cv::Point2f ctr = center(b);
         const float rad =
-            (std::fabs(b.angle) > 45.0f) ? 0.0f : b.angle * static_cast<float>(M_PI) / 180.0f;
+            (std::fabs(b.angle) > 45.0f) ? 0.0f : b.angle * std::numbers::pi_v<float> / 180.0f;
         const auto cs = corners(b);
         for (std::size_t i = 0; i < 4; ++i) {
             q.pts[i] = rotateAround(cs[i], ctr, rad);
@@ -473,15 +479,18 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, floa
             static_cast<float>(minBoxSide + 2)) {
             continue;
         }
-        ::cv::Point2f c[4];
-        expanded.points(c);
+        std::array<::cv::Point2f, 4> c;
+        expanded.points(c.data());
         Quad q;
         q.score = score;
         q.angle = expanded.angle;
-        float minX = static_cast<float>(w), minY = static_cast<float>(h), maxX = 0, maxY = 0;
+        auto minX = static_cast<float>(w);
+        auto minY = static_cast<float>(h);
+        float maxX = 0;
+        float maxY = 0;
         for (int32_t k = 0; k < 4; ++k) {
-            const float px = std::clamp(c[k].x, 0.0f, static_cast<float>(w));
-            const float py = std::clamp(c[k].y, 0.0f, static_cast<float>(h));
+            const float px = std::clamp(c[static_cast<std::size_t>(k)].x, 0.0f, static_cast<float>(w));
+            const float py = std::clamp(c[static_cast<std::size_t>(k)].y, 0.0f, static_cast<float>(h));
             q.pts[static_cast<std::size_t>(k)] = {px, py};
             minX = std::min(minX, px);
             minY = std::min(minY, py);
@@ -499,7 +508,7 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, floa
     // `|dy| > threshold` test is intransitive (a~b, b~c, but a<c) and aborts under
     // libc++ hardening.
     constexpr float kRowBand = 10.0f;
-    std::sort(quads.begin(), quads.end(), [](const Quad &a, const Quad &b) {
+    std::ranges::sort(quads, [](const Quad &a, const Quad &b) {
         const int rowA = static_cast<int>(std::floor(a.pts[0].y / kRowBand));
         const int rowB = static_cast<int>(std::floor(b.pts[0].y / kRowBand));
         if (rowA != rowB) {
@@ -528,7 +537,7 @@ jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
 } // namespace
 
 void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
-    auto name = "extractTextBoxes";
+    const auto *name = "extractTextBoxes";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
                      size_t count) -> jsi::Value {
         if (count != 2) {
@@ -559,7 +568,7 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
             throw jsi::JSError(rt, "extractTextBoxes: src tensor has been disposed");
         }
 
-        const auto *dataPtr = reinterpret_cast<const float *>(src->data_.get());
+        auto *dataPtr = reinterpret_cast<float *>(src->data_.get());
         std::span<const float> data(dataPtr, src->numel_);
 
         std::vector<Quad> quads;
@@ -587,7 +596,7 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                 }
                 const int32_t w = s[s.size() - 1];
                 const int32_t h = s[s.size() - 2];
-                ::cv::Mat prob(h, w, CV_32F, const_cast<float *>(dataPtr));
+                ::cv::Mat prob(h, w, CV_32F, dataPtr);
                 quads = extractDbnet(
                     prob, static_cast<float>(getNumberProp(rt, opts, "binThreshold")),
                     static_cast<float>(getNumberProp(rt, opts, "boxThreshold")),
@@ -610,7 +619,7 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
 
 // --------------------------- ctcGreedyDecode -------------------------------
 void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
-    auto name = "ctcGreedyDecode";
+    const auto *name = "ctcGreedyDecode";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
                      size_t count) -> jsi::Value {
         if (count != 2) {
@@ -656,9 +665,9 @@ void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
         for (int32_t t = 0; t < timesteps; ++t) {
             const float *row = data + static_cast<std::size_t>(t) * static_cast<std::size_t>(vocab);
             const float *maxIt = std::max_element(row, row + vocab);
-            const int32_t maxIdx = static_cast<int32_t>(maxIt - row);
+            const auto maxIdx = static_cast<int32_t>(maxIt - row);
             const float maxVal = *maxIt;
-            double prob = static_cast<double>(maxVal);
+            auto prob = static_cast<double>(maxVal);
             if (softmax) {
                 double sum = 0.0;
                 for (int32_t v = 0; v < vocab; ++v) {

From 98bd0c14c39ec260c29673aee2b1c50198a4247c Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Wed, 1 Jul 2026 19:11:19 +0200
Subject: [PATCH 21/29] =?UTF-8?q?[RNE=20Rewrite]=20refactor(cv):=20reorg?=
 =?UTF-8?q?=20native=20ops=20=E2=80=94=20add=20rotate,=20split=20text=5Fbo?=
 =?UTF-8?q?xes=5Fops,=20inline=20JSI=20option=20readers,=20warpQuad=20offs?=
 =?UTF-8?q?et/clear?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../cpp/core/model.cpp                        |   4 +-
 .../cpp/extensions/cv/image_ops.cpp           | 251 ++++-------------
 .../cpp/extensions/cv/image_ops.h             |   6 +-
 .../cpp/extensions/cv/install.cpp             |   7 +-
 .../cpp/extensions/cv/ocr_ops.cpp             |  30 +-
 .../cpp/extensions/cv/text_boxes_ops.cpp      | 266 ++++++++++++++++++
 .../cpp/extensions/cv/text_boxes_ops.h        |   8 +
 .../cpp/extensions/cv/utils.h                 |  39 ---
 8 files changed, 355 insertions(+), 256 deletions(-)
 create mode 100644 packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
 create mode 100644 packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h

diff --git a/packages/react-native-executorch/cpp/core/model.cpp b/packages/react-native-executorch/cpp/core/model.cpp
index f4b11c4e4b..88d2d418f1 100644
--- a/packages/react-native-executorch/cpp/core/model.cpp
+++ b/packages/react-native-executorch/cpp/core/model.cpp
@@ -455,10 +455,10 @@ jsi::Value ModelHostObject::get(jsi::Runtime &rt, const jsi::PropNameID &name) {
 
             // Free a single previously-executed method's planned-memory activation
             // arena (and, on graph-compiling backends like CoreML, its compiled
-            // graph). The method transparently reloads on next execute. Returns
+            // graph). The method reloads on next execute. Returns
             // whether a loaded method was actually freed (false = not loaded, a
             // harmless no-op). Bounds memory when many distinct bucketed methods
-            // (detect_<S>/recognize_<W>) accumulate over a session.
+            // accumulate over a session.
             auto methodName = args[0].asString(rt).utf8(rt);
             bool unloaded = self->etModule_->unload_method(methodName);
             return jsi::Value(unloaded);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index 31cef5025c..c0e5b513f9 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -716,236 +716,99 @@ void install_applyColormap(jsi::Runtime &rt, jsi::Object &module) {
     module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
 }
 
-void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
-    const auto *name = "gridSample";
-    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
+void install_rotate(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "rotate";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value & /*thisVal*/, const jsi::Value *args, size_t count) -> jsi::Value {
         if (count != 3) {
-            throw jsi::JSError(rt, "Usage: gridSample(src, grid, dst)");
-        }
-        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt) ||
-            !args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt) ||
-            !args[2].isObject() || !args[2].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "gridSample: src, grid, and dst must be Tensors");
+            throw jsi::JSError(rt, "Usage: rotate(src, dst, degCW)");
         }
 
-        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto grid = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto dst = args[2].asObject(rt).getHostObject<TensorHostObject>(rt);
-
-        if (src.get() == dst.get()) {
-            throw jsi::JSError(rt, "gridSample: In-place operations (src == dst) are not supported.");
-        }
-        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
-            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
-            throw jsi::JSError(rt, "gridSample: src and dst must be uint8");
-        }
-        if (grid->dtype_ != rnexecutorch::core::types::DType::float32) {
-            throw jsi::JSError(rt, "gridSample: grid must be float32");
-        }
-        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
-            throw jsi::JSError(rt, "gridSample: src and dst must be [H, W, C]");
-        }
-        if (src->shape_ != dst->shape_) {
-            throw jsi::JSError(rt, "gridSample: src and dst must have the same shape");
-        }
-        // grid is the torch grid_sample field [..,2,gH,gW], channel 0 = x, 1 = y,
-        // normalized to [-1,1] with align_corners=true.
-        const auto &gs = grid->shape_;
-        if (gs.size() < 3 || gs[gs.size() - 3] != 2) {
-            throw jsi::JSError(rt, "gridSample: grid must be [..,2,gH,gW]");
-        }
-
-        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
-        std::shared_lock<std::shared_mutex> gridLock(grid->mutex_, std::try_to_lock);
-        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
-        if (!srcLock.owns_lock() || !gridLock.owns_lock() || !dstLock.owns_lock()) {
-            throw jsi::JSError(rt, "gridSample: a tensor is currently in use");
-        }
-        if (!src->data_ || !grid->data_ || !dst->data_) {
-            throw jsi::JSError(rt, "gridSample: a tensor has been disposed");
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "rotate: src must be a Tensor");
         }
 
-        const int32_t h = src->shape_[0];
-        const int32_t w = src->shape_[1];
-        const int32_t channels = src->shape_[2];
-        const int32_t gridH = gs[gs.size() - 2];
-        const int32_t gridW = gs[gs.size() - 1];
-        const int32_t plane = gridH * gridW;
-        const auto *g = reinterpret_cast<const float *>(grid->data_.get());
-
-        // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
-        auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
-            const int32_t x0 = std::clamp(static_cast<int32_t>(std::floor(gx)), 0, gridW - 1);
-            const int32_t y0 = std::clamp(static_cast<int32_t>(std::floor(gy)), 0, gridH - 1);
-            const int32_t x1 = std::min(x0 + 1, gridW - 1);
-            const int32_t y1 = std::min(y0 + 1, gridH - 1);
-            const float dx = gx - static_cast<float>(x0);
-            const float dy = gy - static_cast<float>(y0);
-            const int32_t base = c * plane;
-            const float top = g[base + y0 * gridW + x0] +
-                              (g[base + y0 * gridW + x1] - g[base + y0 * gridW + x0]) * dx;
-            const float bot = g[base + y1 * gridW + x0] +
-                              (g[base + y1 * gridW + x1] - g[base + y1 * gridW + x0]) * dx;
-            return top + (bot - top) * dy;
-        };
-
-        ::cv::Mat mapX(h, w, CV_32F);
-        ::cv::Mat mapY(h, w, CV_32F);
-        for (int32_t oy = 0; oy < h; ++oy) {
-            const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
-                                         static_cast<float>(gridH - 1)
-                                   : 0.0f;
-            auto *rowX = mapX.ptr<float>(oy);
-            auto *rowY = mapY.ptr<float>(oy);
-            for (int32_t ox = 0; ox < w; ++ox) {
-                const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
-                                             static_cast<float>(gridW - 1)
-                                       : 0.0f;
-                const float nx = sampleGrid(0, gx, gy); // [-1,1]
-                const float ny = sampleGrid(1, gx, gy);
-                rowX[ox] = ((nx + 1.0f) / 2.0f) * static_cast<float>(w - 1);
-                rowY[ox] = ((ny + 1.0f) / 2.0f) * static_cast<float>(h - 1);
-            }
+        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "rotate: dst must be a Tensor");
         }
 
-        const int cvType = CV_MAKETYPE(CV_8U, channels);
-        ::cv::Mat srcMat(h, w, cvType, src->data_.get());
-        ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
-        try {
-            ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
-        } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("gridSample: OpenCV error: ") + e.what());
+        if (!args[2].isNumber()) {
+            throw jsi::JSError(rt, "rotate: degCW must be a number (90, 180, or 270)");
         }
-        return jsi::Value(rt, args[2]);
-    };
-    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
-}
 
-// ------------------------------- warpQuad ----------------------------------
-// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
-// resize-to-height + pad/align). A generic image op; used by the OCR recognizer.
-void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
-    const auto *name = "warpQuad";
-    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
-                     size_t count) -> jsi::Value {
-        if (count != 4) {
-            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
-        }
-        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
-        }
-        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
-        }
-        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
-            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
-        }
-        if (!args[3].isObject()) {
-            throw jsi::JSError(rt, "warpQuad: options must be an object");
-        }
         auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
         auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+
         if (src.get() == dst.get()) {
-            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
+            throw jsi::JSError(rt, "rotate: In-place operations (src == dst) are not supported.");
         }
-        auto quadArr = args[2].asObject(rt).asArray(rt);
-        auto opts = args[3].asObject(rt);
 
-        if (quadArr.length(rt) != 8) {
-            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
+        const auto degCW = static_cast<int32_t>(args[2].asNumber());
+        int rotateCode = 0;
+        if (degCW == 90) {
+            rotateCode = ::cv::ROTATE_90_CLOCKWISE;
+        } else if (degCW == 180) {
+            rotateCode = ::cv::ROTATE_180;
+        } else if (degCW == 270) {
+            rotateCode = ::cv::ROTATE_90_COUNTERCLOCKWISE;
+        } else {
+            throw jsi::JSError(rt, "rotate: degCW must be 90, 180, or 270");
         }
+
         if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
+            throw jsi::JSError(rt, "rotate: src and dst must be [H, W, C]");
         }
-        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
-            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
+
+        if (src->dtype_ != dst->dtype_) {
+            throw jsi::JSError(rt, "rotate: src and dst must have the same dtype");
         }
+
         if (src->shape_[2] != dst->shape_[2]) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
+            throw jsi::JSError(rt, "rotate: src and dst must have the same number of channels");
         }
 
+        const int32_t srcH = src->shape_[0];
+        const int32_t srcW = src->shape_[1];
         const int32_t channels = src->shape_[2];
-        const int32_t recH = dst->shape_[0];
-        const int32_t bucketW = dst->shape_[1];
-
-        if (!opts.hasProperty(rt, "contentWidth") ||
-            !opts.getProperty(rt, "contentWidth").isNumber()) {
-            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
-        }
-        const int32_t contentWidth =
-            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
-                       bucketW);
-        const std::string padMode = getStringProp(rt, opts, "padMode");
-        const double padValue = getNumberProp(rt, opts, "padValue");
-        const std::string align = getStringProp(rt, opts, "align");
-
-        std::array<::cv::Point2f, 4> quad;
-        for (std::size_t i = 0; i < 8; ++i) {
-            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
-                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
-            }
-        }
-        for (std::size_t i = 0; i < 4; ++i) {
-            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
-                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
+        // 90/270 transpose the axes; 180 preserves them. dst must be pre-sized to match,
+        // else cv::rotate reallocates off the tensor buffer and the result is lost.
+        const bool swap = degCW != 180;
+        const int32_t expH = swap ? srcW : srcH;
+        const int32_t expW = swap ? srcH : srcW;
+        if (dst->shape_[0] != expH || dst->shape_[1] != expW) {
+            throw jsi::JSError(rt, "rotate: dst must be sized [" + std::to_string(expH) + ", " +
+                                       std::to_string(expW) + ", C] for a " + std::to_string(degCW) +
+                                       " degree rotation");
         }
 
         std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
         if (!srcLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
+            throw jsi::JSError(rt, "rotate: src tensor is currently in use");
         }
+
         std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
         if (!dstLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
+            throw jsi::JSError(rt, "rotate: dst tensor is currently in use");
         }
+
         if (!src->data_ || !dst->data_) {
-            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
+            throw jsi::JSError(rt, "rotate: a tensor has been disposed");
         }
 
-        const int cvType = CV_MAKETYPE(CV_8U, channels);
-        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
-        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
-
+        int cvType{};
         try {
-            const std::array<::cv::Point2f, 4> dstPts = {
-                ::cv::Point2f{0.0f, 0.0f},
-                {static_cast<float>(contentWidth), 0.0f},
-                {static_cast<float>(contentWidth), static_cast<float>(recH)},
-                {0.0f, static_cast<float>(recH)}};
-            const std::array<::cv::Point2f, 4> srcPts = {quad[0], quad[1], quad[2], quad[3]};
-            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts.data(), dstPts.data());
-            ::cv::Mat content;
-            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
-                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
-
-            ::cv::Scalar padColor;
-            if (padMode == "cornerMean") {
-                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
-                ::cv::Scalar acc(0, 0, 0, 0);
-                const std::array<::cv::Rect, 4> rects = {
-                    ::cv::Rect(0, 0, patch, patch),
-                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
-                    ::cv::Rect(0, recH - patch, patch, patch),
-                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
-                for (const auto &r : rects) {
-                    acc += ::cv::mean(content(r));
-                }
-                padColor = acc / 4.0;
-            } else {
-                padColor = ::cv::Scalar::all(padValue);
-            }
-
-            dstMat.setTo(padColor);
-            const int32_t offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
-            content.copyTo(dstMat(::cv::Rect(offsetX, 0, contentWidth, recH)));
-        } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
+            cvType = CV_MAKETYPE(dtypeToCvDepth(src->dtype_), channels);
+        } catch (const std::invalid_argument &e) {
+            throw jsi::JSError(rt, "rotate: " + std::string(e.what()));
         }
+
+        const ::cv::Mat srcMat(srcH, srcW, cvType, src->data_.get());
+        ::cv::Mat dstMat(expH, expW, cvType, dst->data_.get());
+        ::cv::rotate(srcMat, dstMat, rotateCode);
+
         return jsi::Value(rt, args[1]);
     };
-    module.setProperty(rt, name,
-                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
-                                                             4, fnBody));
+
+    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
 }
+
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
index 90542dd212..8c91c16878 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
@@ -9,9 +9,5 @@ void install_toChannelsFirst(facebook::jsi::Runtime &rt, facebook::jsi::Object &
 void install_toChannelsLast(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_normalize(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_applyColormap(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-void install_gridSample(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
-// resize-to-height + pad/align), HWC uint8. Generic image op; the OCR recognizer
-// uses it to lift each text quad into a fixed-height recognizer canvas.
-void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+void install_rotate(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/install.cpp b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
index 7377f0d3ce..a228a55654 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/install.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
@@ -2,6 +2,7 @@
 #include "box_ops.h"
 #include "image_ops.h"
 #include "ocr_ops.h"
+#include "text_boxes_ops.h"
 
 namespace rnexecutorch::extensions::cv {
 namespace jsi = facebook::jsi;
@@ -15,8 +16,7 @@ void install(facebook::jsi::Runtime &rt, facebook::jsi::Object &module) {
     image_ops::install_toChannelsLast(rt, cvModule);
     image_ops::install_normalize(rt, cvModule);
     image_ops::install_applyColormap(rt, cvModule);
-    image_ops::install_gridSample(rt, cvModule);
-    image_ops::install_warpQuad(rt, cvModule);
+    image_ops::install_rotate(rt, cvModule);
 
     box_ops::install_nms(rt, cvModule);
     box_ops::install_restrictToBox(rt, cvModule);
@@ -24,6 +24,9 @@ void install(facebook::jsi::Runtime &rt, facebook::jsi::Object &module) {
     ocr_ops::install_extractTextBoxes(rt, cvModule);
     ocr_ops::install_ctcGreedyDecode(rt, cvModule);
 
+    text_boxes_ops::install_gridSample(rt, cvModule);
+    text_boxes_ops::install_warpQuad(rt, cvModule);
+
     module.setProperty(rt, "cv", cvModule);
 }
 } // namespace rnexecutorch::extensions::cv
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 77401e6c23..68f13c545d 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -21,7 +21,6 @@
 
 #include "core/dtype.h"
 #include "core/tensor.h"
-#include "utils.h"
 
 // Detector postprocessing geometry: CRAFT text-map grouping + DBNet prob-map ->
 // oriented quads. Pure OpenCV, kept native. ctcGreedyDecode (per-timestep argmax
@@ -581,13 +580,16 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                 }
                 const int32_t heatW = s[s.size() - 2];
                 const int32_t heatH = s[s.size() - 3];
-                const double targetH = getNumberProp(rt, opts, "targetHeight");
+                const double targetH = opts.getProperty(rt, "targetHeight").asNumber();
                 const float restoreRatio = static_cast<float>(targetH) / static_cast<float>(heatH);
-                quads = extractCraft(data, heatW, heatH,
-                                     static_cast<float>(getNumberProp(rt, opts, "textThreshold")),
-                                     static_cast<float>(getNumberProp(rt, opts, "linkThreshold")),
-                                     static_cast<float>(getNumberProp(rt, opts, "lowTextThreshold")),
-                                     restoreRatio, getBoolPropOr(rt, opts, "charLevel", false));
+                const bool charLevel =
+                    opts.hasProperty(rt, "charLevel") && opts.getProperty(rt, "charLevel").asBool();
+                quads = extractCraft(
+                    data, heatW, heatH,
+                    static_cast<float>(opts.getProperty(rt, "textThreshold").asNumber()),
+                    static_cast<float>(opts.getProperty(rt, "linkThreshold").asNumber()),
+                    static_cast<float>(opts.getProperty(rt, "lowTextThreshold").asNumber()), restoreRatio,
+                    charLevel);
             } else if (mode == "dbnet") {
                 // src is [1,1,H,W] or [H,W] probability map (full-res).
                 const auto &s = src->shape_;
@@ -598,12 +600,12 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                 const int32_t h = s[s.size() - 2];
                 ::cv::Mat prob(h, w, CV_32F, dataPtr);
                 quads = extractDbnet(
-                    prob, static_cast<float>(getNumberProp(rt, opts, "binThreshold")),
-                    static_cast<float>(getNumberProp(rt, opts, "boxThreshold")),
-                    static_cast<float>(getNumberProp(rt, opts, "unclipRatio")),
-                    static_cast<int32_t>(getNumberProp(rt, opts, "minBoxSide")),
-                    static_cast<int32_t>(getNumberProp(rt, opts, "maxCandidates")),
-                    getBoolProp(rt, opts, "applySigmoid"));
+                    prob, static_cast<float>(opts.getProperty(rt, "binThreshold").asNumber()),
+                    static_cast<float>(opts.getProperty(rt, "boxThreshold").asNumber()),
+                    static_cast<float>(opts.getProperty(rt, "unclipRatio").asNumber()),
+                    static_cast<int32_t>(opts.getProperty(rt, "minBoxSide").asNumber()),
+                    static_cast<int32_t>(opts.getProperty(rt, "maxCandidates").asNumber()),
+                    opts.getProperty(rt, "applySigmoid").asBool());
             } else {
                 throw jsi::JSError(rt, "extractTextBoxes: unknown mode '" + mode + "'");
             }
@@ -649,7 +651,7 @@ void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
             throw jsi::JSError(rt, "ctcGreedyDecode: numel must be a multiple of the vocab dim");
         }
         const int32_t timesteps = static_cast<int32_t>(src->numel_) / vocab;
-        const bool softmax = getBoolProp(rt, opts, "softmax");
+        const bool softmax = opts.getProperty(rt, "softmax").asBool();
 
         std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
         if (!srcLock.owns_lock()) {
diff --git a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
new file mode 100644
index 0000000000..59f1e45ff6
--- /dev/null
+++ b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
@@ -0,0 +1,266 @@
+#include "text_boxes_ops.h"
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <numeric>
+#include <stdexcept>
+#include <utility>
+
+#include <opencv2/imgproc.hpp>
+
+#include "core/dtype.h"
+#include "core/tensor.h"
+#include "utils.h"
+namespace rnexecutorch::extensions::cv::text_boxes_ops {
+
+namespace jsi = facebook::jsi;
+using TensorHostObject = rnexecutorch::core::tensor::TensorHostObject;
+
+void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "gridSample";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
+        if (count != 3) {
+            throw jsi::JSError(rt, "Usage: gridSample(src, grid, dst)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[2].isObject() || !args[2].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "gridSample: src, grid, and dst must be Tensors");
+        }
+
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto grid = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[2].asObject(rt).getHostObject<TensorHostObject>(rt);
+
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "gridSample: In-place operations (src == dst) are not supported.");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "gridSample: src and dst must be uint8");
+        }
+        if (grid->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "gridSample: grid must be float32");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "gridSample: src and dst must be [H, W, C]");
+        }
+        if (src->shape_ != dst->shape_) {
+            throw jsi::JSError(rt, "gridSample: src and dst must have the same shape");
+        }
+        // grid is the torch grid_sample field [..,2,gH,gW], channel 0 = x, 1 = y,
+        // normalized to [-1,1] with align_corners=true.
+        const auto &gs = grid->shape_;
+        if (gs.size() < 3 || gs[gs.size() - 3] != 2) {
+            throw jsi::JSError(rt, "gridSample: grid must be [..,2,gH,gW]");
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        std::shared_lock<std::shared_mutex> gridLock(grid->mutex_, std::try_to_lock);
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock() || !gridLock.owns_lock() || !dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "gridSample: a tensor is currently in use");
+        }
+        if (!src->data_ || !grid->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "gridSample: a tensor has been disposed");
+        }
+
+        const int32_t h = src->shape_[0];
+        const int32_t w = src->shape_[1];
+        const int32_t channels = src->shape_[2];
+        const int32_t gridH = gs[gs.size() - 2];
+        const int32_t gridW = gs[gs.size() - 1];
+        const int32_t plane = gridH * gridW;
+        const auto *g = reinterpret_cast<const float *>(grid->data_.get());
+
+        // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
+        auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
+            const int32_t x0 = std::clamp(static_cast<int32_t>(std::floor(gx)), 0, gridW - 1);
+            const int32_t y0 = std::clamp(static_cast<int32_t>(std::floor(gy)), 0, gridH - 1);
+            const int32_t x1 = std::min(x0 + 1, gridW - 1);
+            const int32_t y1 = std::min(y0 + 1, gridH - 1);
+            const float dx = gx - static_cast<float>(x0);
+            const float dy = gy - static_cast<float>(y0);
+            const int32_t base = c * plane;
+            const float top = g[base + y0 * gridW + x0] +
+                              (g[base + y0 * gridW + x1] - g[base + y0 * gridW + x0]) * dx;
+            const float bot = g[base + y1 * gridW + x0] +
+                              (g[base + y1 * gridW + x1] - g[base + y1 * gridW + x0]) * dx;
+            return top + (bot - top) * dy;
+        };
+
+        ::cv::Mat mapX(h, w, CV_32F);
+        ::cv::Mat mapY(h, w, CV_32F);
+        for (int32_t oy = 0; oy < h; ++oy) {
+            const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
+                                         static_cast<float>(gridH - 1)
+                                   : 0.0f;
+            auto *rowX = mapX.ptr<float>(oy);
+            auto *rowY = mapY.ptr<float>(oy);
+            for (int32_t ox = 0; ox < w; ++ox) {
+                const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
+                                             static_cast<float>(gridW - 1)
+                                       : 0.0f;
+                const float nx = sampleGrid(0, gx, gy); // [-1,1]
+                const float ny = sampleGrid(1, gx, gy);
+                rowX[ox] = ((nx + 1.0f) / 2.0f) * static_cast<float>(w - 1);
+                rowY[ox] = ((ny + 1.0f) / 2.0f) * static_cast<float>(h - 1);
+            }
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(h, w, cvType, src->data_.get());
+        ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
+        try {
+            ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("gridSample: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[2]);
+    };
+    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
+}
+
+// ------------------------------- warpQuad ----------------------------------
+// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
+// resize-to-height + pad/align). A generic image op; used by the OCR recognizer.
+void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "warpQuad";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 4) {
+            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
+        }
+        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
+        }
+        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
+            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
+        }
+        if (!args[3].isObject()) {
+            throw jsi::JSError(rt, "warpQuad: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
+        }
+        auto quadArr = args[2].asObject(rt).asArray(rt);
+        auto opts = args[3].asObject(rt);
+
+        if (quadArr.length(rt) != 8) {
+            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
+        }
+        if (src->shape_[2] != dst->shape_[2]) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
+        }
+
+        const int32_t channels = src->shape_[2];
+        const int32_t recH = dst->shape_[0];
+        const int32_t bucketW = dst->shape_[1];
+
+        if (!opts.hasProperty(rt, "contentWidth") ||
+            !opts.getProperty(rt, "contentWidth").isNumber()) {
+            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
+        }
+        const int32_t contentWidth =
+            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
+                       bucketW);
+        const std::string padMode = opts.getProperty(rt, "padMode").asString(rt).utf8(rt);
+        const double padValue = opts.getProperty(rt, "padValue").asNumber();
+        const std::string align = opts.getProperty(rt, "align").asString(rt).utf8(rt);
+        // offsetX >= 0 places content at that x (overriding align); clear=false skips
+        // wiping dst first, so successive warps compose into one canvas (glyph strips).
+        const auto offsetXOpt = static_cast<int32_t>(opts.getProperty(rt, "offsetX").asNumber());
+        const bool clear = opts.getProperty(rt, "clear").asBool();
+
+        std::array<::cv::Point2f, 4> quad;
+        for (std::size_t i = 0; i < 8; ++i) {
+            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
+                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
+            }
+        }
+        for (std::size_t i = 0; i < 4; ++i) {
+            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
+                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
+        }
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
+        }
+        if (!src->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
+        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
+
+        try {
+            const std::array<::cv::Point2f, 4> dstPts = {
+                ::cv::Point2f{0.0f, 0.0f},
+                {static_cast<float>(contentWidth), 0.0f},
+                {static_cast<float>(contentWidth), static_cast<float>(recH)},
+                {0.0f, static_cast<float>(recH)}};
+            const std::array<::cv::Point2f, 4> srcPts = {quad[0], quad[1], quad[2], quad[3]};
+            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts.data(), dstPts.data());
+            ::cv::Mat content;
+            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
+                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
+
+            ::cv::Scalar padColor;
+            if (padMode == "cornerMean") {
+                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
+                ::cv::Scalar acc(0, 0, 0, 0);
+                const std::array<::cv::Rect, 4> rects = {
+                    ::cv::Rect(0, 0, patch, patch),
+                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
+                    ::cv::Rect(0, recH - patch, patch, patch),
+                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
+                for (const auto &r : rects) {
+                    acc += ::cv::mean(content(r));
+                }
+                padColor = acc / 4.0;
+            } else {
+                padColor = ::cv::Scalar::all(padValue);
+            }
+
+            if (clear) {
+                dstMat.setTo(padColor);
+            }
+            int32_t offsetX = offsetXOpt;
+            if (offsetX < 0) {
+                offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
+            }
+            if (offsetX < bucketW) {
+                const int32_t copyW = std::min(contentWidth, bucketW - offsetX);
+                content(::cv::Rect(0, 0, copyW, recH))
+                    .copyTo(dstMat(::cv::Rect(offsetX, 0, copyW, recH)));
+            }
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[1]);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             4, fnBody));
+}
+
+} // namespace rnexecutorch::extensions::cv::text_boxes_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h
new file mode 100644
index 0000000000..99ad684e28
--- /dev/null
+++ b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h
@@ -0,0 +1,8 @@
+#pragma once
+
+#include <jsi/jsi.h>
+
+namespace rnexecutorch::extensions::cv::text_boxes_ops {
+void install_gridSample(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+} // namespace rnexecutorch::extensions::cv::text_boxes_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/utils.h b/packages/react-native-executorch/cpp/extensions/cv/utils.h
index 75a9de194a..bb6a42ecac 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/utils.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/utils.h
@@ -20,43 +20,4 @@ inline int dtypeToCvDepth(rnexecutorch::core::types::DType dtype) {
     throw std::invalid_argument("unsupported dtype");
 }
 
-// ------------------------------ JSI option readers ------------------------------
-// Required option getters shared by the cv op installers. Defaults live in the TS
-// wrappers, so a missing/mis-typed property is a contract violation -> throw.
-inline double getNumberProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
-                            const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isNumber()) {
-        throw facebook::jsi::JSError(rt, std::string("options.") + name +
-                                             " is required and must be a number");
-    }
-    return opts.getProperty(rt, name).asNumber();
-}
-
-inline std::string getStringProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
-                                 const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isString()) {
-        throw facebook::jsi::JSError(rt, std::string("options.") + name +
-                                             " is required and must be a string");
-    }
-    return opts.getProperty(rt, name).asString(rt).utf8(rt);
-}
-
-inline bool getBoolProp(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
-                        const char *name) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
-        throw facebook::jsi::JSError(rt, std::string("options.") + name +
-                                             " is required and must be a boolean");
-    }
-    return opts.getProperty(rt, name).asBool();
-}
-
-// Optional boolean (defaults when absent) — used for flags a caller may omit.
-inline bool getBoolPropOr(facebook::jsi::Runtime &rt, const facebook::jsi::Object &opts,
-                          const char *name, bool fallback) {
-    if (!opts.hasProperty(rt, name) || !opts.getProperty(rt, name).isBool()) {
-        return fallback;
-    }
-    return opts.getProperty(rt, name).asBool();
-}
-
 } // namespace rnexecutorch::extensions::cv

From 3640b17f0cb9578a465c2358a37ee98786583d0d Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Wed, 1 Jul 2026 19:11:37 +0200
Subject: [PATCH 22/29] [RNE Rewrite] refactor(ocr): move into cv/ocr, split
 textBoxes ops, tensor-thread document pipeline + Category-B builders

---
 .../src/extensions/cv/ops/image.ts            | 193 +-------------
 .../src/extensions/cv/ops/points.ts           |  13 -
 .../src/extensions/cv/ops/textBoxes.ts        | 206 +++++++++++++++
 .../src/extensions/cv/tasks/documentOCR.ts    |  82 ++++--
 .../src/extensions/cv/tasks/ocr.ts            |  93 +++++--
 .../cv/tasks/{ => ocr}/documentHelpers.ts     |  54 +---
 .../cv/tasks/{ => ocr}/ocrHelpers.ts          |   8 +-
 .../cv/tasks/{ => ocr}/ocrPipeline.ts         | 250 ++++++++----------
 .../cv/tasks/ocr}/ocrSymbols.ts               |   0
 .../cv/tasks/{ => ocr}/supporting.ts          |  85 +++---
 .../src/extensions/cv/tasks/preprocessing.ts  |  36 ++-
 11 files changed, 544 insertions(+), 476 deletions(-)
 create mode 100644 packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/documentHelpers.ts (75%)
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/ocrHelpers.ts (98%)
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/ocrPipeline.ts (72%)
 rename packages/react-native-executorch/src/{ => extensions/cv/tasks/ocr}/ocrSymbols.ts (100%)
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/supporting.ts (79%)

diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index ed30273f4c..c7eec8fd8b 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -1,7 +1,6 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ImageFormat } from '../image';
-import type { Point } from './points';
 
 /**
  * Supported color conversion code presets (similar to OpenCV).
@@ -212,192 +211,16 @@ export function applyColormap(
 }
 
 /**
- * Detector heatmap postprocessing mode. `'craft'` consumes a half-resolution
- * interleaved `[..,Hd,Wd,2]` (text, affinity) map; `'dbnet'` consumes a
- * full-resolution `[..,H,W]` probability map.
- * @category Types
- */
-export type TextBoxMode = 'craft' | 'dbnet';
-
-/**
- * An oriented quadrilateral text region returned by {@link extractTextBoxes},
- * in detector-input pixel space.
- * @category Types
- */
-export type Quad = {
-  /** Four corner points, ordered TL, TR, BR, BL. */
-  readonly points: readonly Point[];
-  /** Detector confidence for this region (1 for CRAFT grouped lines). */
-  readonly score: number;
-  /** Line angle in degrees. */
-  readonly angle: number;
-};
-
-/**
- * CRAFT-mode options for {@link extractTextBoxes} (region+affinity heatmap).
- * @category Types
- */
-export type CraftExtractOptions = {
-  readonly mode: 'craft';
-  readonly textThreshold: number;
-  readonly linkThreshold: number;
-  readonly lowTextThreshold: number;
-  /**
-   * Detector input height — used to restore the half-res boxes. Required (the
-   * native op throws without it); the OCR pipeline sets it per run.
-   */
-  readonly targetHeight: number;
-  /**
-   * Return one upright box per glyph instead of grouped reading-ordered lines:
-   * the affinity map is subtracted (not added) to separate adjacent characters,
-   * and no line grouping or de-skew rotation is applied. Used by the second,
-   * per-column pass that reads upright stacked text. Defaults to `false`.
-   */
-  readonly charLevel?: boolean;
-};
-
-/**
- * DBNet-mode options for {@link extractTextBoxes} (probability map).
- * @category Types
- */
-export type DbnetExtractOptions = {
-  readonly mode: 'dbnet';
-  readonly binThreshold: number;
-  readonly boxThreshold: number;
-  readonly unclipRatio: number;
-  readonly minBoxSide: number;
-  readonly maxCandidates: number;
-  /**
-   * Whether the probability map needs a sigmoid applied first. Set from the
-   * model's export contract (`true` = raw logits, `false` = already probabilities).
-   */
-  readonly applySigmoid: boolean;
-};
-
-/**
- * Options for {@link extractTextBoxes} — a discriminated union on `mode`, so the
- * type system enforces exactly the field set the chosen detector needs.
- * @category Types
- */
-export type ExtractTextBoxesOptions = CraftExtractOptions | DbnetExtractOptions;
-
-/**
- * Converts a detector heatmap tensor into oriented text-line quads.
- *
- * Variable-sized output: the native op returns a flat array of 10 doubles per
- * box which this wrapper reshapes into {@link Quad}s.
- * @category Typescript API
- * @param src The detector output tensor (float32). CRAFT: `[..,Hd,Wd,2]`;
- * DBNet: `[..,H,W]`.
- * @param opts Mode and per-mode thresholds.
- * @returns The detected quads in detector-input pixel space.
- */
-export function extractTextBoxes(src: Tensor, opts: ExtractTextBoxesOptions): Quad[] {
-  'worklet';
-  const flat = rnexecutorchJsi.cv.extractTextBoxes(src, opts) as number[];
-  const quads: Quad[] = [];
-  for (let i = 0; i < flat.length; i += 10) {
-    quads.push({
-      points: [
-        { x: flat[i]!, y: flat[i + 1]! },
-        { x: flat[i + 2]!, y: flat[i + 3]! },
-        { x: flat[i + 4]!, y: flat[i + 5]! },
-        { x: flat[i + 6]!, y: flat[i + 7]! },
-      ],
-      score: flat[i + 8]!,
-      angle: flat[i + 9]!,
-    });
-  }
-  return quads;
-}
-
-/**
- * Options for {@link warpQuad}.
- * @category Types
- */
-export type WarpQuadOptions = {
-  /** Width (px) of the warped content inside the destination canvas. */
-  readonly contentWidth: number;
-  /** Horizontal placement of the content in the canvas. Default `'left'`. */
-  readonly align?: 'left' | 'center';
-  /** How the remaining canvas is filled. Default `'constant'`. */
-  readonly padMode?: 'constant' | 'cornerMean';
-  /** Fill value for `padMode: 'constant'`. Default `0`. */
-  readonly padValue?: number;
-};
-
-/**
- * Perspective-crops an oriented quad region of `src` into the pre-allocated
- * recognizer canvas `dst`, folding crop + resize-to-height + bucket-pad.
- * @category Typescript API
- * @param src The source image tensor in HWC uint8 layout. Shape [H,W,C].
- * @param dst The pre-allocated recognizer canvas in HWC uint8 layout, sized
- * `[recH, bucketW, C]`.
- * @param quad Eight numbers `[x0,y0,..,x3,y3]` (TL,TR,BR,BL) in `src` pixels.
- * @param opts Content width, alignment, and padding configuration.
- * @returns The destination tensor `dst`.
- */
-export function warpQuad(src: Tensor, dst: Tensor, quad: number[], opts: WarpQuadOptions): Tensor {
-  'worklet';
-  // Defaults live here (the native op requires every field) — see add-native-extension.
-  return rnexecutorchJsi.cv.warpQuad(src, dst, quad, {
-    contentWidth: opts.contentWidth,
-    align: opts.align ?? 'left',
-    padMode: opts.padMode ?? 'constant',
-    padValue: opts.padValue ?? 0,
-  });
-}
-
-/**
- * Options for {@link ctcGreedyDecode}.
- * @category Types
- */
-export type CtcGreedyDecodeOptions = {
-  /**
-   * Apply a per-timestep softmax so `values` are probabilities (CRNN logits).
-   * Leave `false` for heads that are already softmaxed (SVTR).
-   */
-  readonly softmax?: boolean;
-};
-
-/**
- * Per-timestep argmax + max value over recognizer logits, computed natively on
- * the tensor buffer (avoids copying the whole `[T, V]` tensor into JS).
- * @category Typescript API
- * @param src The recognizer output tensor (float32), shape `[..,T,V]`.
- * @param opts Whether to softmax each timestep before taking the max value.
- * @returns The per-timestep argmax `indices` and (optionally softmaxed) max
- * `values`.
- */
-export function ctcGreedyDecode(
-  src: Tensor,
-  opts?: CtcGreedyDecodeOptions
-): { indices: number[]; values: number[] } {
-  'worklet';
-  const flat = rnexecutorchJsi.cv.ctcGreedyDecode(src, {
-    softmax: opts?.softmax ?? false,
-  }) as number[];
-  const indices: number[] = [];
-  const values: number[] = [];
-  for (let i = 0; i < flat.length; i += 2) {
-    indices.push(flat[i]!);
-    values.push(flat[i + 1]!);
-  }
-  return { indices, values };
-}
-
-/**
- * Resamples `src` through a backward sampling field (the `torch.grid_sample`
- * step of a geometric dewarp) into the pre-allocated `dst`, natively via
- * `cv::remap`.
+ * Rotates `src` clockwise by `degCW` degrees (90, 180, or 270) into the
+ * pre-allocated `dst`. A 90/270 rotation swaps width and height, so `dst` must be
+ * sized with `src`'s height and width transposed.
  * @category Typescript API
- * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
- * @param grid The sampling field tensor (float32), shape `[..,2,gH,gW]`, channel
- * 0 = x and 1 = y, normalized to `[-1, 1]` with `align_corners=true`.
- * @param dst The pre-allocated destination tensor, same shape/dtype as `src`.
+ * @param src The source image tensor (HWC).
+ * @param dst The destination tensor, pre-sized for the rotation.
+ * @param degCW The clockwise rotation in degrees: 90, 180, or 270.
  * @returns The destination tensor `dst`.
  */
-export function gridSample(src: Tensor, grid: Tensor, dst: Tensor): Tensor {
+export function rotate(src: Tensor, dst: Tensor, degCW: number): Tensor {
   'worklet';
-  return rnexecutorchJsi.cv.gridSample(src, grid, dst);
+  return rnexecutorchJsi.cv.rotate(src, dst, degCW);
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index e7d8f7c71d..47cbeedfa4 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -9,19 +9,6 @@ export type Point = {
   readonly y: number;
 };
 
-/**
- * Clamps a scalar to the inclusive range `[lo, hi]`.
- * @category Utils
- * @param v The value to clamp.
- * @param lo The lower bound.
- * @param hi The upper bound.
- * @returns `v` constrained to `[lo, hi]`.
- */
-export function clamp(v: number, lo: number, hi: number): number {
-  'worklet';
-  return v < lo ? lo : v > hi ? hi : v;
-}
-
 /**
  * Per-axis scale and offset that map a `to`-space coordinate back into
  * `from`-space, for an aspect-preserving letterbox or an axis stretch. The
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
new file mode 100644
index 0000000000..429047f930
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
@@ -0,0 +1,206 @@
+import { rnexecutorchJsi } from '../../../native/bridge';
+import type { Tensor } from '../../../core/tensor';
+import type { Point } from './points';
+
+/**
+ * Detector heatmap postprocessing mode. `'craft'` consumes a half-resolution
+ * interleaved `[..,Hd,Wd,2]` (text, affinity) map; `'dbnet'` consumes a
+ * full-resolution `[..,H,W]` probability map.
+ * @category Types
+ */
+export type TextBoxMode = 'craft' | 'dbnet';
+
+/**
+ * An oriented quadrilateral text region returned by {@link extractTextBoxes},
+ * in detector-input pixel space.
+ * @category Types
+ */
+export type Quad = {
+  /** Four corner points, ordered TL, TR, BR, BL. */
+  readonly points: readonly Point[];
+  /** Detector confidence for this region (1 for CRAFT grouped lines). */
+  readonly score: number;
+  /** Line angle in degrees. */
+  readonly angle: number;
+};
+
+/**
+ * CRAFT-mode options for {@link extractTextBoxes} (region+affinity heatmap).
+ * @category Types
+ */
+export type CraftExtractOptions = {
+  readonly mode: 'craft';
+  readonly textThreshold: number;
+  readonly linkThreshold: number;
+  readonly lowTextThreshold: number;
+  /**
+   * Detector input height — used to restore the half-res boxes. Required (the
+   * native op throws without it); the OCR pipeline sets it per run.
+   */
+  readonly targetHeight: number;
+  /**
+   * Return one upright box per glyph instead of grouped reading-ordered lines:
+   * the affinity map is subtracted (not added) to separate adjacent characters,
+   * and no line grouping or de-skew rotation is applied. Used by the second,
+   * per-column pass that reads upright stacked text. Defaults to `false`.
+   */
+  readonly charLevel?: boolean;
+};
+
+/**
+ * DBNet-mode options for {@link extractTextBoxes} (probability map).
+ * @category Types
+ */
+export type DbnetExtractOptions = {
+  readonly mode: 'dbnet';
+  readonly binThreshold: number;
+  readonly boxThreshold: number;
+  readonly unclipRatio: number;
+  readonly minBoxSide: number;
+  readonly maxCandidates: number;
+  /**
+   * Whether the probability map needs a sigmoid applied first. Set from the
+   * model's export contract (`true` = raw logits, `false` = already probabilities).
+   */
+  readonly applySigmoid: boolean;
+};
+
+/**
+ * Options for {@link extractTextBoxes} — a discriminated union on `mode`, so the
+ * type system enforces exactly the field set the chosen detector needs.
+ * @category Types
+ */
+export type ExtractTextBoxesOptions = CraftExtractOptions | DbnetExtractOptions;
+
+/**
+ * Converts a detector heatmap tensor into oriented text-line quads.
+ *
+ * Variable-sized output: the native op returns a flat array of 10 doubles per
+ * box which this wrapper reshapes into {@link Quad}s.
+ * @category Typescript API
+ * @param src The detector output tensor (float32). CRAFT: `[..,Hd,Wd,2]`;
+ * DBNet: `[..,H,W]`.
+ * @param opts Mode and per-mode thresholds.
+ * @returns The detected quads in detector-input pixel space.
+ */
+export function extractTextBoxes(src: Tensor, opts: ExtractTextBoxesOptions): Quad[] {
+  'worklet';
+  const flat = rnexecutorchJsi.cv.extractTextBoxes(src, opts) as number[];
+  const quads: Quad[] = [];
+  for (let i = 0; i < flat.length; i += 10) {
+    quads.push({
+      points: [
+        { x: flat[i]!, y: flat[i + 1]! },
+        { x: flat[i + 2]!, y: flat[i + 3]! },
+        { x: flat[i + 4]!, y: flat[i + 5]! },
+        { x: flat[i + 6]!, y: flat[i + 7]! },
+      ],
+      score: flat[i + 8]!,
+      angle: flat[i + 9]!,
+    });
+  }
+  return quads;
+}
+
+/**
+ * Options for {@link warpQuad}.
+ * @category Types
+ */
+export type WarpQuadOptions = {
+  /** Width (px) of the warped content inside the destination canvas. */
+  readonly contentWidth: number;
+  /** Horizontal placement of the content in the canvas. Default `'left'`. */
+  readonly align?: 'left' | 'center';
+  /** How the remaining canvas is filled. Default `'constant'`. */
+  readonly padMode?: 'constant' | 'cornerMean';
+  /** Fill value for `padMode: 'constant'`. Default `0`. */
+  readonly padValue?: number;
+  /**
+   * Exact x (px) to place the content at, overriding `align`. Default `-1`
+   * (use `align`). Combined with `clear: false`, lets successive warps compose
+   * side-by-side into one canvas (e.g. a glyph strip).
+   */
+  readonly offsetX?: number;
+  /**
+   * Wipe the canvas to the pad color before writing. Default `true`. Set `false`
+   * to preserve prior content (compose multiple warps into one `dst`).
+   */
+  readonly clear?: boolean;
+};
+
+/**
+ * Perspective-crops an oriented quad region of `src` into the pre-allocated
+ * recognizer canvas `dst`, folding crop + resize-to-height + bucket-pad.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout. Shape [H,W,C].
+ * @param dst The pre-allocated recognizer canvas in HWC uint8 layout, sized
+ * `[recH, bucketW, C]`.
+ * @param quad Eight numbers `[x0,y0,..,x3,y3]` (TL,TR,BR,BL) in `src` pixels.
+ * @param opts Content width, alignment, and padding configuration.
+ * @returns The destination tensor `dst`.
+ */
+export function warpQuad(src: Tensor, dst: Tensor, quad: number[], opts: WarpQuadOptions): Tensor {
+  'worklet';
+  return rnexecutorchJsi.cv.warpQuad(src, dst, quad, {
+    contentWidth: opts.contentWidth,
+    align: opts.align ?? 'left',
+    padMode: opts.padMode ?? 'constant',
+    padValue: opts.padValue ?? 0,
+    offsetX: opts.offsetX ?? -1,
+    clear: opts.clear ?? true,
+  });
+}
+
+/**
+ * Options for {@link ctcGreedyDecode}.
+ * @category Types
+ */
+export type CtcGreedyDecodeOptions = {
+  /**
+   * Apply a per-timestep softmax so `values` are probabilities (CRNN logits).
+   * Leave `false` for heads that are already softmaxed (SVTR).
+   */
+  readonly softmax?: boolean;
+};
+
+/**
+ * Per-timestep argmax + max value over recognizer logits, computed natively on
+ * the tensor buffer (avoids copying the whole `[T, V]` tensor into JS).
+ * @category Typescript API
+ * @param src The recognizer output tensor (float32), shape `[..,T,V]`.
+ * @param opts Whether to softmax each timestep before taking the max value.
+ * @returns The per-timestep argmax `indices` and (optionally softmaxed) max
+ * `values`.
+ */
+export function ctcGreedyDecode(
+  src: Tensor,
+  opts?: CtcGreedyDecodeOptions
+): { indices: number[]; values: number[] } {
+  'worklet';
+  const flat = rnexecutorchJsi.cv.ctcGreedyDecode(src, {
+    softmax: opts?.softmax ?? false,
+  }) as number[];
+  const indices: number[] = [];
+  const values: number[] = [];
+  for (let i = 0; i < flat.length; i += 2) {
+    indices.push(flat[i]!);
+    values.push(flat[i + 1]!);
+  }
+  return { indices, values };
+}
+
+/**
+ * Resamples `src` through a backward sampling field (the `torch.grid_sample`
+ * step of a geometric dewarp) into the pre-allocated `dst`, natively via
+ * `cv::remap`.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
+ * @param grid The sampling field tensor (float32), shape `[..,2,gH,gW]`, channel
+ * 0 = x and 1 = y, normalized to `[-1, 1]` with `align_corners=true`.
+ * @param dst The pre-allocated destination tensor, same shape/dtype as `src`.
+ * @returns The destination tensor `dst`.
+ */
+export function gridSample(src: Tensor, grid: Tensor, dst: Tensor): Tensor {
+  'worklet';
+  return rnexecutorchJsi.cv.gridSample(src, grid, dst);
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
index 7de026f093..12d0bb98df 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
@@ -1,18 +1,20 @@
 import type { WorkletRuntime } from 'react-native-worklets';
 
+import { tensor } from '../../../core/tensor';
 import { wrapAsync } from '../../../core/runtime';
 import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
 import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
+import { rotate, FORMAT_CHANNELS } from '../ops/image';
 import { createOCR, type OCRModel, type OCRDetection } from './ocr';
 import {
   createObjectDetector,
   type ObjectDetectorModel,
   type ObjectDetection,
 } from './objectDetection';
-import { createSupporting, type SupportingModel } from './supporting';
-import { readingOrderIndices } from './ocrHelpers';
-import { cropImageBuffer, rotateImageBuffer, fillTableCells } from './documentHelpers';
+import { createSupporting, type SupportingModel } from './ocr/supporting';
+import { readingOrderIndices } from './ocr/ocrHelpers';
+import { cropImageBuffer, fillTableCells } from './ocr/documentHelpers';
 
 /**
  * One assembled document block: a layout region (or an ungrouped catch-all) with
@@ -98,11 +100,6 @@ export type RunDocumentOCROptions = {
 // can still land ~0.74.
 const ORIENTATION_MIN_CONFIDENCE = 0.85;
 
-const isTableLabel = (label: unknown): boolean => {
-  'worklet';
-  return String(label) === 'table';
-};
-
 // Layout classes that carry no text — skip OCR on them.
 const VISUAL_LABELS = ['image', 'chart', 'seal'];
 const isTextRegion = (label: unknown): boolean => {
@@ -186,18 +183,54 @@ export async function createDocumentOCR<L>(
     const useOrientation = !!supporting && (options?.orientation ?? defaultOrientation);
     const useDewarp = !!supporting && (options?.dewarp ?? defaultDewarp);
     let img = input;
-    if (useOrientation && supporting) {
-      // Only correct when the classifier is confident AND the predicted angle is
-      // non-zero — a low-confidence argmax (typical of OOD photos / non-documents)
-      // otherwise spuriously flips the page. Below threshold we leave it as 0°.
-      const ori = supporting.detectOrientationWorklet(img);
-      if (ori.rotationCW !== 0 && ori.confidence >= ORIENTATION_MIN_CONFIDENCE) {
-        img = rotateImageBuffer(img, ((360 - ori.rotationCW) % 360) as 0 | 90 | 180 | 270);
+    // Orientation + dewarp thread one page tensor (built once): rotate and the
+    // dewarp remap run tensor -> tensor with no intermediate ImageBuffer, and the
+    // corrected page is materialized back to an ImageBuffer once for the OCR/crop
+    // stages below (which are all ImageBuffer-based).
+    if ((useOrientation || useDewarp) && supporting) {
+      const ch = FORMAT_CHANNELS[input.format];
+      let page = tensor('uint8', [input.height, input.width, ch]);
+      page.setData(input.data);
+      let pw = input.width;
+      let ph = input.height;
+      try {
+        if (useOrientation) {
+          // Only correct when the classifier is confident AND the predicted angle
+          // is non-zero — a low-confidence argmax (typical of OOD photos / non-
+          // documents) otherwise spuriously flips the page.
+          const orientation = supporting.detectOrientationWorklet(page, input.format);
+          const deg = ((360 - orientation.rotationCW) % 360) as 0 | 90 | 180 | 270;
+          if (deg !== 0 && orientation.confidence >= ORIENTATION_MIN_CONFIDENCE) {
+            const swap = deg === 90 || deg === 270;
+            const rotated = tensor('uint8', [swap ? pw : ph, swap ? ph : pw, ch]);
+            try {
+              rotate(page, rotated, deg);
+            } catch (e) {
+              rotated.dispose(); // rotate threw before we adopted `rotated` as `page`
+              throw e;
+            }
+            page.dispose();
+            page = rotated;
+            if (swap) {
+              [pw, ph] = [ph, pw];
+            }
+          }
+        }
+        if (useDewarp) {
+          // dewarp returns the input tensor unchanged when it declines the warp.
+          const dewarped = supporting.dewarpWorklet(page, input.format);
+          if (dewarped !== page) {
+            page.dispose();
+            page = dewarped;
+          }
+        }
+        const out = new Uint8Array(pw * ph * ch);
+        page.getData(out);
+        img = { data: out, width: pw, height: ph, format: input.format, layout: input.layout };
+      } finally {
+        page.dispose();
       }
     }
-    if (useDewarp && supporting) {
-      img = supporting.dewarpWorklet(img);
-    }
 
     // OCR runs once per region here (potentially many), so don't let each call
     // free+reload its bucket arenas (release: false). Instead free the model's
@@ -236,13 +269,18 @@ export async function createDocumentOCR<L>(
         const lines = ocr
           .runOCRWorklet(crop, { release: false })
           .detections.map((d) => offsetDetection(d, xmin, ymin));
-        const table = isTableLabel(region.label);
-        if (lines.length === 0 && !table) {
+        if (lines.length === 0 && region.label !== 'table') {
           continue;
         }
         detections.push(...lines);
-        let block = makeBlock<L>(region.label, region.box, region.confidence, lines, table);
-        if (table && supporting) {
+        let block = makeBlock<L>(
+          region.label,
+          region.box,
+          region.confidence,
+          lines,
+          region.label === 'table'
+        );
+        if (region.label === 'table' && supporting) {
           const structure = supporting.recognizeTableWorklet(crop);
           block = { ...block, tableHtml: fillTableCells(structure.html, block.lines) };
         }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index e417e5182f..26498c6833 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -6,7 +6,8 @@ import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
-import { FORMAT_CHANNELS, cvtColor, type Quad } from '../ops/image';
+import { FORMAT_CHANNELS, cvtColor } from '../ops/image';
+import type { Quad } from '../ops/textBoxes';
 import {
   buildCharset,
   orderQuad,
@@ -16,7 +17,7 @@ import {
   boundingQuadOf,
   groupVerticalColumns,
   type Buckets,
-} from './ocrHelpers';
+} from './ocr/ocrHelpers';
 import {
   toRgbCode,
   detectQuads,
@@ -25,8 +26,8 @@ import {
   readBoxVertical,
   validateDetectorSchema,
   buildExtractOpts,
-  buildDetectorSets,
-  buildRecognizerSets,
+  deriveDetectorOutputs,
+  deriveRecognizerContract,
   disposeDetSets,
   disposeRecSets,
   type DetSet,
@@ -35,9 +36,9 @@ import {
   type RecContext,
   type VerticalContext,
   type DetectorExtractConfig,
-} from './ocrPipeline';
+} from './ocr/ocrPipeline';
 
-export type { Buckets } from './ocrHelpers';
+export type { Buckets } from './ocr/ocrHelpers';
 
 /**
  * Configuration for the unified OCR pipeline. A model declares its detector
@@ -282,23 +283,51 @@ export async function createOCR(
       ocrOpts.extractBoxes
     );
 
-    const built = buildRecognizerSets(model, recBuckets);
-    recSets = built.sets;
-    recC = built.recC;
-    recH = built.recH;
-    recSetByWidth = new Map(recSets.map((s) => [s.width, s]));
+    // Derive shapes/contract, then allocate + own the scratch tensors here so
+    // ownership never crosses a function boundary (the derive* helpers return no
+    // tensors).
+    const rec = deriveRecognizerContract(model, recBuckets);
+    recC = rec.recC;
+    recH = rec.recH;
+    // Push into the pre-declared arrays as we allocate, so a mid-loop tensor()
+    // failure leaves the partial set visible to the catch's dispose* below.
+    for (const bucket of rec.buckets) {
+      recSets.push({
+        width: bucket.width,
+        tCanvas: tensor('uint8', [rec.recH, bucket.width, rec.recC]),
+        tCF: tensor('uint8', [rec.recC, rec.recH, bucket.width]),
+        tNorm: tensor('float32', [rec.recC, rec.recH, bucket.width]),
+        tInput: tensor('float32', bucket.inShape),
+        tLogits: tensor('float32', bucket.outShape),
+      });
+    }
+    recSetByWidth = new Map(recSets.map((recSet) => [recSet.width, recSet]));
 
     if (recC !== 3) {
       throw new Error(`OCR: recognizer must take RGB (3 channels), but the model expects ${recC}.`);
     }
     charset = buildCharset(ocrOpts.charset);
-    if (charset.length !== built.vocabSize) {
+    if (charset.length !== rec.vocabSize) {
       throw new Error(
-        `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${built.vocabSize}).`
+        `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${rec.vocabSize}).`
       );
     }
-    detSets = buildDetectorSets(model, detBuckets, ocrOpts.detectorKind, detExtraChannels);
-    detSetByS = new Map(detSets.map((d) => [d.s, d]));
+    for (const { s, outputs } of deriveDetectorOutputs(
+      model,
+      detBuckets,
+      ocrOpts.detectorKind,
+      detExtraChannels
+    )) {
+      detSets.push({
+        s,
+        tColor: tensor('uint8', [s, s, 3]),
+        tCF: tensor('uint8', [3, s, s]),
+        tNorm: tensor('float32', [3, s, s]),
+        tInput: tensor('float32', [1, 3, s, s]),
+        tOutputs: outputs.map((o) => tensor(o.dtype, o.shape)),
+      });
+    }
+    detSetByS = new Map(detSets.map((detSet) => [detSet.s, detSet]));
   } catch (e) {
     disposeRecSets(recSets);
     disposeDetSets(detSets);
@@ -391,18 +420,18 @@ export async function createOCR(
       // Valid (non-tiny) boxes, ordered TL,TR,BR,BL.
       const ordered: Point[][] = [];
       for (const quad of quads) {
-        const o = orderQuad(quad);
-        const s = quadSize(o);
-        if (s.width >= 3 && s.height >= 3) {
-          ordered.push(o);
+        const orderedQuad = orderQuad(quad);
+        const size = quadSize(orderedQuad);
+        if (size.width >= 3 && size.height >= 3) {
+          ordered.push(orderedQuad);
         }
       }
 
       if (!vertical) {
-        for (const o of ordered) {
+        for (const orderedQuad of ordered) {
           const boxStart = nowMs();
-          const { text, conf } = recognizeQuad(recCtx, recSrc, o);
-          pushDetection(detections, dropScore, text, conf, o, nowMs() - boxStart);
+          const { text, conf } = recognizeQuad(recCtx, recSrc, orderedQuad);
+          pushDetection(detections, dropScore, text, conf, orderedQuad, nowMs() - boxStart);
         }
         return { detections: orderDetections(detections) };
       }
@@ -415,27 +444,33 @@ export async function createOCR(
         // `recognizeGlyphStrip` splits any multi-letter box into single-glyph
         // cells (DBNet merges stacked letters and won't split them), so the
         // column's boxes can be passed straight through, top -> bottom.
-        const r = recognizeGlyphStrip(recCtx, recSrc, recC, col);
-        if (r) {
+        const strip = recognizeGlyphStrip(recCtx, recSrc, col);
+        if (strip) {
           pushDetection(
             detections,
             VERTICAL_DROP_SCORE,
-            r.text,
-            r.conf,
+            strip.text,
+            strip.conf,
             boundingQuadOf(col),
             nowMs() - boxStart
           );
         }
       }
-      for (const o of singles) {
+      for (const orderedQuad of singles) {
         const boxStart = nowMs();
-        const { text, conf, stacked } = readBoxVertical(recCtx, vctx, recSrc, o, quadSize(o));
+        const { text, conf, stacked } = readBoxVertical(
+          recCtx,
+          vctx,
+          recSrc,
+          orderedQuad,
+          quadSize(orderedQuad)
+        );
         pushDetection(
           detections,
           stacked ? VERTICAL_DROP_SCORE : dropScore,
           text,
           conf,
-          o,
+          orderedQuad,
           nowMs() - boxStart
         );
       }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
similarity index 75%
rename from packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
index c35d40a434..8d7f8cbbcd 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
@@ -1,8 +1,7 @@
-import type { ImageBuffer } from '../image';
-import { FORMAT_CHANNELS } from '../ops/image';
-import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
-import { clamp } from '../ops/points';
-import type { OCRDetection } from './ocr';
+import type { ImageBuffer } from '../../image';
+import { FORMAT_CHANNELS } from '../../ops/image';
+import { boundingBoxOf, type BoundingBox } from '../../ops/boxes';
+import type { OCRDetection } from '../ocr';
 
 // Crops an axis-aligned region out of an ImageBuffer (pure pixel slice, same
 // format). Used to feed a layout region (e.g. a table) to another model.
@@ -10,10 +9,10 @@ export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>):
   'worklet';
   const { data, width, height, format } = input;
   const ch = FORMAT_CHANNELS[format];
-  const x0 = clamp(Math.round(bbox.xmin), 0, width);
-  const y0 = clamp(Math.round(bbox.ymin), 0, height);
-  const x1 = clamp(Math.round(bbox.xmax), 0, width);
-  const y1 = clamp(Math.round(bbox.ymax), 0, height);
+  const x0 = Math.max(0, Math.min(Math.round(bbox.xmin), width));
+  const y0 = Math.max(0, Math.min(Math.round(bbox.ymin), height));
+  const x1 = Math.max(0, Math.min(Math.round(bbox.xmax), width));
+  const y1 = Math.max(0, Math.min(Math.round(bbox.ymax), height));
   const cw = Math.max(1, x1 - x0);
   const chh = Math.max(1, y1 - y0);
   const out = new Uint8Array(cw * chh * ch);
@@ -24,43 +23,6 @@ export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>):
   return { data: out, width: cw, height: chh, format, layout: input.layout };
 }
 
-// Rotates an ImageBuffer clockwise by 0/90/180/270 degrees (pure pixel reindex).
-export function rotateImageBuffer(input: ImageBuffer, degCW: 0 | 90 | 180 | 270): ImageBuffer {
-  'worklet';
-  if (degCW === 0) {
-    return input;
-  }
-  const { data, width, height, format } = input;
-  const ch = FORMAT_CHANNELS[format];
-  const swap = degCW === 90 || degCW === 270;
-  const ow = swap ? height : width;
-  const oh = swap ? width : height;
-  const out = new Uint8Array(ow * oh * ch);
-  for (let y = 0; y < height; y++) {
-    for (let x = 0; x < width; x++) {
-      let ox = 0;
-      let oy = 0;
-      if (degCW === 90) {
-        ox = height - 1 - y;
-        oy = x;
-      } else if (degCW === 180) {
-        ox = width - 1 - x;
-        oy = height - 1 - y;
-      } else {
-        // 270
-        ox = y;
-        oy = width - 1 - x;
-      }
-      const si = (y * width + x) * ch;
-      const di = (oy * ow + ox) * ch;
-      for (let c = 0; c < ch; c++) {
-        out[di + c] = data[si + c]!;
-      }
-    }
-  }
-  return { data: out, width: ow, height: oh, format, layout: input.layout };
-}
-
 /**
  * Index of the maximum value in `arr[offset, offset+len)` (single pass, no
  * allocation). Shared by the layout / orientation / table decoders.
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
similarity index 98%
rename from packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
index 32555903c1..fc467231af 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
@@ -1,5 +1,5 @@
-import type { Quad } from '../ops/image';
-import { scalePoint, clamp, type Point } from '../ops/points';
+import type { Quad } from '../../ops/textBoxes';
+import { scalePoint, type Point } from '../../ops/points';
 
 /**
  * Worklet-safe millisecond clock for in-pipeline profiling. Prefers
@@ -94,7 +94,7 @@ export function mapQuadToImage(
       to: { width: origW, height: origH },
       resizeMode: 'letterbox',
     });
-    return { x: clamp(m.x, 0, origW), y: clamp(m.y, 0, origH) };
+    return { x: Math.max(0, Math.min(m.x, origW)), y: Math.max(0, Math.min(m.y, origH)) };
   });
 }
 
@@ -454,7 +454,7 @@ export function contentWidthFor(
 ): number {
   'worklet';
   const w = Math.round((recHeight * quadW) / Math.max(1, quadH));
-  return clamp(w, 1, bucketWidth);
+  return Math.max(1, Math.min(w, bucketWidth));
 }
 
 /**
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
similarity index 72%
rename from packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
index 4111db4057..20745b7ac5 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocrPipeline.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
@@ -9,26 +9,28 @@
 // readStackedColumn/recognizeQuad). The non-worklet builders/validators run at
 // construction time on the JS thread and have no such constraint.
 
-import { tensor, type Tensor } from '../../../core/tensor';
-import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
-import type { Model } from '../../../core/model';
+import { tensor, type Tensor, type DType } from '../../../../core/tensor';
+import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
+import type { Model } from '../../../../core/model';
 
-import type { ImageFormat } from '../image';
-import type { Point } from '../ops/points';
+import type { ImageFormat } from '../../image';
+import type { Point } from '../../ops/points';
 import {
   FORMAT_CONVERSION,
   resize,
   cvtColor,
   toChannelsFirst,
   normalize,
+  type ColorConversionCode,
+} from '../../ops/image';
+import {
   extractTextBoxes,
   warpQuad,
   ctcGreedyDecode,
-  type ColorConversionCode,
   type CraftExtractOptions,
   type DbnetExtractOptions,
   type Quad,
-} from '../ops/image';
+} from '../../ops/textBoxes';
 import {
   mapQuadToImage,
   orderQuad,
@@ -156,25 +158,25 @@ export function detectQuads(
   'worklet';
   const detS = snapDetectBucket(width, height, ctx.detBuckets);
   // snapDetectBucket always returns one of detBuckets, so the set exists.
-  const ds = ctx.detSets.get(detS)!;
+  const detSet = ctx.detSets.get(detS)!;
   // Only the source resize depends on the run's channel count; the rest is cached.
   const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
   try {
     src
       .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
-      .throughIf(ctx.detCode !== null, cvtColor, ds.tColor, ctx.detCode!)
-      .through(toChannelsFirst, ds.tCF)
-      .through(normalize, ds.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
-      .copyTo(ds.tInput);
+      .throughIf(ctx.detCode !== null, cvtColor, detSet.tColor, ctx.detCode!)
+      .through(toChannelsFirst, detSet.tCF)
+      .through(normalize, detSet.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
+      .copyTo(detSet.tInput);
 
-    ctx.model.execute(`detect_${detS}`, [ds.tInput], [...ds.tOutputs]);
+    ctx.model.execute(`detect_${detS}`, [detSet.tInput], [...detSet.tOutputs]);
     // A custom arch hands its raw outputs to the user extractor; the built-ins
     // decode the heatmap (tOutputs[0]). CRAFT needs the per-run input height to
     // restore its half-res boxes; `charLevel` switches to per-glyph boxes.
     const quads = ctx.extractBoxes
-      ? ctx.extractBoxes(ds.tOutputs, detS)
+      ? ctx.extractBoxes(detSet.tOutputs, detS)
       : extractTextBoxes(
-          ds.tOutputs[0]!,
+          detSet.tOutputs[0]!,
           ctx.extractOpts!.mode === 'craft'
             ? { ...ctx.extractOpts!, targetHeight: detS, charLevel }
             : ctx.extractOpts!
@@ -198,121 +200,108 @@ export function recognizeQuad(
   const desiredW = contentWidthFor(cs.width, cs.height, ctx.recH, maxRec);
   const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
   // snapRecognizeBucket always returns one of recBuckets, so the set exists.
-  const rs = ctx.recSetByWidth.get(bucketW)!;
+  const recSet = ctx.recSetByWidth.get(bucketW)!;
   const contentWidth = Math.min(desiredW, bucketW);
-  warpQuad(src, rs.tCanvas, flattenQuad(corners), {
+  warpQuad(src, recSet.tCanvas, flattenQuad(corners), {
     contentWidth,
     align: 'left',
     padMode: 'constant',
     padValue: ctx.padValue,
   });
-  rs.tCanvas
-    .through(toChannelsFirst, rs.tCF)
-    .through(normalize, rs.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
-    .copyTo(rs.tInput);
-  ctx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
+  recSet.tCanvas
+    .through(toChannelsFirst, recSet.tCF)
+    .through(normalize, recSet.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
+    .copyTo(recSet.tInput);
+  ctx.model.execute(`recognize_${bucketW}`, [recSet.tInput], [recSet.tLogits]);
   // A custom decode (e.g. attention/AR head) takes the raw logits; otherwise
   // greedy CTC. Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
   if (ctx.decode) {
-    const r = ctx.decode(rs.tLogits, ctx.charset);
+    const r = ctx.decode(recSet.tLogits, ctx.charset);
     return { text: r.text, conf: r.confidence };
   }
-  const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
+  const { indices, values } = ctcGreedyDecode(recSet.tLogits, { softmax: false });
   const text = decodeGreedy(indices, ctx.charset);
   const conf = ctcConfidence(values, indices);
   return { text, conf };
 }
 
-// Joins glyph quads (in `src` pixel space, in reading order) into one recognizer
-// strip — each glyph warped upright to the recognizer height and laid side by
-// side — and recognizes it as a single line (joint hconcat). Returns null when
-// nothing usable was assembled.
+// Joins glyph quads (in `src` pixel space, reading order) into one recognizer
+// strip: each glyph is warped upright to the recognizer height and placed side by
+// side directly in the canvas (native `warpQuad` with a per-glyph `offsetX`, so
+// there is no JS pixel assembly), then recognized as a single line. Returns null
+// when nothing usable was assembled.
 //
 // Must be defined BEFORE its callers: the worklet plugin captures referenced
 // worklets in source order, so a forward reference is undefined at run time.
 export function recognizeGlyphStrip(
   recCtx: RecContext,
   src: Tensor,
-  recC: number,
   glyphs: readonly (readonly Point[])[]
 ): { text: string; conf: number } | null {
   'worklet';
   const recH = recCtx.recH;
   const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-  // Warp each glyph upright to recognizer height (aspect preserved). A box that
-  // is much taller than wide is a merged run of stacked letters — split it into
-  // ~square single-letter cells first, so each lands in its own strip slot
-  // (otherwise N letters get squashed into one cell and read as garbage).
-  const slices: { tGlyph: Tensor; w: number }[] = [];
+  // Pass 1 (geometry only): a box much taller than wide is a merged run of stacked
+  // letters — split it into ~square single-letter cells so each lands in its own
+  // strip slot. Measure each cell's warped width (aspect preserved) to size the strip.
+  const cells: { quad: readonly Point[]; width: number }[] = [];
   let totalW = 0;
-  for (const g of glyphs) {
-    const gsz = quadSize(g);
-    if (gsz.width < 1 || gsz.height < 1) {
+  for (const glyph of glyphs) {
+    const glyphSize = quadSize(glyph);
+    if (glyphSize.width < 1 || glyphSize.height < 1) {
       continue;
     }
-    const parts = Math.max(1, Math.round(gsz.height / Math.max(1, gsz.width)));
-    for (const cell of splitTallQuad(g, parts)) {
-      const gs = quadSize(cell);
-      if (gs.width < 1 || gs.height < 1) {
+    const parts = Math.max(1, Math.round(glyphSize.height / Math.max(1, glyphSize.width)));
+    for (const cell of splitTallQuad(glyph, parts)) {
+      const cellSize = quadSize(cell);
+      if (cellSize.width < 1 || cellSize.height < 1) {
         continue;
       }
-      const gw = Math.max(1, Math.min(Math.round((gs.width * recH) / gs.height), maxRec));
-      const tGlyph = tensor('uint8', [recH, gw, recC]);
-      warpQuad(src, tGlyph, flattenQuad(cell), {
-        contentWidth: gw,
-        align: 'left',
-        padMode: 'constant',
-        padValue: recCtx.padValue,
-      });
-      slices.push({ tGlyph, w: gw });
-      totalW += gw;
+      const width = Math.max(
+        1,
+        Math.min(Math.round((cellSize.width * recH) / cellSize.height), maxRec)
+      );
+      cells.push({ quad: cell, width });
+      totalW += width;
     }
   }
-  if (slices.length === 0) {
+  if (cells.length === 0) {
     return null;
   }
-  try {
-    // Smallest bucket that fits the strip (snap up, no glyph truncated); widest
-    // bucket for very long columns.
-    const bucketW =
-      recCtx.recBuckets.find((w) => w >= totalW) ??
-      recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-    const rs = recCtx.recSetByWidth.get(bucketW)!;
-    // Assemble the strip row-major into the bucket canvas, neutral-padded.
-    const strip = new Uint8Array(recH * bucketW * recC);
-    strip.fill(recCtx.padValue);
-    let xOff = 0;
-    for (const s of slices) {
-      if (xOff >= bucketW) {
-        break;
-      }
-      const copyW = Math.min(s.w, bucketW - xOff);
-      const glyphBytes = new Uint8Array(recH * s.w * recC);
-      s.tGlyph.getData(glyphBytes);
-      for (let oy = 0; oy < recH; oy++) {
-        const srcStart = oy * s.w * recC;
-        const row = glyphBytes.subarray(srcStart, srcStart + copyW * recC);
-        strip.set(row, (oy * bucketW + xOff) * recC);
-      }
-      xOff += s.w;
+  // Smallest bucket that fits the strip (snap up, no glyph truncated); widest for
+  // very long columns.
+  const bucketW =
+    recCtx.recBuckets.find((w) => w >= totalW) ?? recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+  const recSet = recCtx.recSetByWidth.get(bucketW)!;
+  // Pass 2: warp each cell straight into the canvas at its x-offset. The first warp
+  // clears + pads the whole canvas; the rest compose in with `clear: false`.
+  let xOff = 0;
+  for (let i = 0; i < cells.length; i++) {
+    if (xOff >= bucketW) {
+      break;
     }
-    rs.tCanvas.setData(strip);
-    rs.tCanvas
-      .through(toChannelsFirst, rs.tCF)
-      .through(normalize, rs.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
-      .copyTo(rs.tInput);
-    recCtx.model.execute(`recognize_${bucketW}`, [rs.tInput], [rs.tLogits]);
-    if (recCtx.decode) {
-      const r = recCtx.decode(rs.tLogits, recCtx.charset);
-      return r.text.length > 0 ? { text: r.text, conf: r.confidence } : null;
-    }
-    const { indices, values } = ctcGreedyDecode(rs.tLogits, { softmax: false });
-    const text = decodeGreedy(indices, recCtx.charset);
-    const conf = ctcConfidence(values, indices);
-    return text.length > 0 ? { text, conf } : null;
-  } finally {
-    slices.forEach((s) => s.tGlyph.dispose());
+    warpQuad(src, recSet.tCanvas, flattenQuad(cells[i]!.quad), {
+      contentWidth: cells[i]!.width,
+      offsetX: xOff,
+      clear: i === 0,
+      padMode: 'constant',
+      padValue: recCtx.padValue,
+    });
+    xOff += cells[i]!.width;
+  }
+  recSet.tCanvas
+    .through(toChannelsFirst, recSet.tCF)
+    .through(normalize, recSet.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
+    .copyTo(recSet.tInput);
+  recCtx.model.execute(`recognize_${bucketW}`, [recSet.tInput], [recSet.tLogits]);
+  if (recCtx.decode) {
+    const decoded = recCtx.decode(recSet.tLogits, recCtx.charset);
+    return decoded.text.length > 0 ? { text: decoded.text, conf: decoded.confidence } : null;
   }
+  const { indices, values } = ctcGreedyDecode(recSet.tLogits, { softmax: false });
+  const text = decodeGreedy(indices, recCtx.charset);
+  const conf = ctcConfidence(values, indices);
+  return text.length > 0 ? { text, conf } : null;
 }
 
 // Reads a single tall box that packs several stacked glyphs the detector grouped
@@ -352,7 +341,7 @@ export function readStackedColumn(
     const boxSrc = vctx.recCode !== null ? cvtColor(tBoxRaw, tRecBox, vctx.recCode) : tBoxRaw;
     // Stack reading order: top -> bottom by each glyph's upper edge.
     const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
-    return recognizeGlyphStrip(recCtx, boxSrc, recC, glyphs);
+    return recognizeGlyphStrip(recCtx, boxSrc, glyphs);
   } finally {
     tBoxRaw.dispose();
     tRecBox.dispose();
@@ -458,53 +447,54 @@ export function buildExtractOpts(
       };
 }
 
-// Pre-allocates one detector scratch-set per detect bucket (channel-independent
-// tensors; the per-run source-resize tensor is allocated in detectQuads). Mirrors
-// buildRecognizerSets — runs at construction.
-export function buildDetectorSets(
+// Per-detect-bucket output tensor specs (dtype + shape). Custom archs declare
+// arbitrary outputs — read straight from the PTE's method metadata; the built-ins
+// have known heatmap (+ craft extras) shapes. Returns specs only (no tensors); the
+// task factory allocates and owns them (see DetSet in createOCR). Runs at
+// construction.
+export function deriveDetectorOutputs(
   model: Model,
   detBuckets: readonly number[],
   detectorKind: 'craft' | 'dbnet' | 'custom',
   detExtraChannels: readonly number[]
-): DetSet[] {
+): { s: number; outputs: { dtype: DType; shape: number[] }[] }[] {
   return detBuckets.map((s) => {
     const heat = s / 2;
-    // Custom archs declare arbitrary outputs — size them straight from the PTE's
-    // method metadata. Built-ins keep their known heatmap (+ craft extras) shapes.
-    let tOutputs: Tensor[];
     if (detectorKind === 'custom') {
-      tOutputs = model
+      const outputs = model
         .getMethodMeta(`detect_${s}`)
-        .outputTensorMeta.map((m) => tensor(m.dtype, m.shape));
-    } else {
-      const tHeatmap =
-        detectorKind === 'dbnet'
-          ? tensor('float32', [1, 1, s, s])
-          : tensor('float32', [1, heat, heat, 2]);
-      tOutputs = [tHeatmap, ...detExtraChannels.map((c) => tensor('float32', [1, c, heat, heat]))];
+        .outputTensorMeta.map((m) => ({ dtype: m.dtype, shape: m.shape }));
+      return { s, outputs };
     }
-    return {
-      s,
-      tColor: tensor('uint8', [s, s, 3]),
-      tCF: tensor('uint8', [3, s, s]),
-      tNorm: tensor('float32', [3, s, s]),
-      tInput: tensor('float32', [1, 3, s, s]),
-      tOutputs,
-    };
+    const heatmap: { dtype: DType; shape: number[] } =
+      detectorKind === 'dbnet'
+        ? { dtype: 'float32', shape: [1, 1, s, s] }
+        : { dtype: 'float32', shape: [1, heat, heat, 2] };
+    const outputs = [
+      heatmap,
+      ...detExtraChannels.map((c) => ({ dtype: 'float32' as DType, shape: [1, c, heat, heat] })),
+    ];
+    return { s, outputs };
   });
 }
 
-// Pre-allocates one recognizer tensor-set per width bucket (each `recognize_<W>`
-// validated once) and derives the constant channel/height/vocab contract from the
-// first bucket. Kept out of the task factory; runs at construction.
-export function buildRecognizerSets(
+// Validates each `recognize_<W>` method and derives the recognizer contract:
+// the constant channel/height/vocab (from the first bucket) plus each bucket's
+// input/output shapes. Returns specs only (no tensors); the task factory allocates
+// and owns the RecSet tensors. Runs at construction.
+export function deriveRecognizerContract(
   model: Model,
   recBuckets: readonly number[]
-): { sets: RecSet[]; recC: number; recH: number; vocabSize: number } {
+): {
+  recC: number;
+  recH: number;
+  vocabSize: number;
+  buckets: { width: number; inShape: number[]; outShape: number[] }[];
+} {
   let recC = 0;
   let recH = 0;
   let vocabSize = 0;
-  const sets = recBuckets.map((w, i) => {
+  const buckets = recBuckets.map((w, i) => {
     const m = validateModelSchema(
       model,
       `recognize_${w}`,
@@ -512,22 +502,16 @@ export function buildRecognizerSets(
       [SymbolicTensor('float32', [1, 'T', 'V'])]
     );
     const inShape = m.inputTensorMeta[0]!.shape;
+    const outShape = m.outputTensorMeta[0]!.shape;
     if (i === 0) {
       // Channels/height/vocab are constant across the width buckets.
       recC = inShape[1]!;
       recH = inShape[2]!;
-      vocabSize = m.outputTensorMeta[0]!.shape[2]!;
+      vocabSize = outShape[2]!;
     }
-    return {
-      width: w,
-      tCanvas: tensor('uint8', [recH, w, recC]),
-      tCF: tensor('uint8', [recC, recH, w]),
-      tNorm: tensor('float32', [recC, recH, w]),
-      tInput: tensor('float32', inShape),
-      tLogits: tensor('float32', m.outputTensorMeta[0]!.shape),
-    };
+    return { width: w, inShape, outShape };
   });
-  return { sets, recC, recH, vocabSize };
+  return { recC, recH, vocabSize, buckets };
 }
 
 // Frees a detector scratch-set's tensors (input prep + per-bucket outputs).
diff --git a/packages/react-native-executorch/src/ocrSymbols.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrSymbols.ts
similarity index 100%
rename from packages/react-native-executorch/src/ocrSymbols.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrSymbols.ts
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
similarity index 79%
rename from packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
index 8c70702ba7..330806d794 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/supporting.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
@@ -1,14 +1,15 @@
 import type { WorkletRuntime } from 'react-native-worklets';
 
-import { tensor } from '../../../core/tensor';
-import { loadModel } from '../../../core/model';
-import { validateModelSchema, SymbolicTensor } from '../../../core/modelSchema';
-import { wrapAsync } from '../../../core/runtime';
+import { tensor, type Tensor } from '../../../../core/tensor';
+import { loadModel } from '../../../../core/model';
+import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
+import { wrapAsync } from '../../../../core/runtime';
 
-import type { ImageBuffer } from '../image';
-import { IMAGENET_NORM } from '../../../constants';
-import { gridSample, FORMAT_CHANNELS } from '../ops/image';
-import { createImagePreprocessor } from './preprocessing';
+import type { ImageBuffer, ImageFormat } from '../../image';
+import { IMAGENET_NORM } from '../../../../constants';
+import { FORMAT_CHANNELS } from '../../ops/image';
+import { gridSample } from '../../ops/textBoxes';
+import { createImagePreprocessor } from '../preprocessing';
 import { argmaxRange } from './documentHelpers';
 
 // SLANet defaults; overridable per model via SupportingModel (the vocab's eos
@@ -108,10 +109,10 @@ export async function createSupporting(
   runtime?: WorkletRuntime
 ): Promise<{
   dispose: () => void;
-  detectOrientation: (input: ImageBuffer) => Promise<Orientation>;
-  detectOrientationWorklet: (input: ImageBuffer) => Orientation;
-  dewarp: (input: ImageBuffer) => Promise<ImageBuffer>;
-  dewarpWorklet: (input: ImageBuffer) => ImageBuffer;
+  detectOrientation: (page: Tensor, format: ImageFormat) => Promise<Orientation>;
+  detectOrientationWorklet: (page: Tensor, format: ImageFormat) => Orientation;
+  dewarp: (page: Tensor, format: ImageFormat) => Promise<Tensor>;
+  dewarpWorklet: (page: Tensor, format: ImageFormat) => Tensor;
   recognizeTable: (input: ImageBuffer) => Promise<TableStructure>;
   recognizeTableWorklet: (input: ImageBuffer) => TableStructure;
 }> {
@@ -169,7 +170,7 @@ export async function createSupporting(
     );
   }
 
-  const oriPre = createImagePreprocessor(
+  const orientationPreprocessor = createImagePreprocessor(
     {
       resizeMode: 'stretch',
       interpolation: 'linear',
@@ -178,11 +179,11 @@ export async function createSupporting(
     },
     oriShape
   );
-  const dewPre = createImagePreprocessor(
+  const dewarpPreprocessor = createImagePreprocessor(
     { resizeMode: 'stretch', interpolation: 'linear', alpha: 1 / 255, beta: 0 },
     dewShape
   );
-  const tabPre = createImagePreprocessor(
+  const tablePreprocessor = createImagePreprocessor(
     {
       resizeMode: 'stretch',
       interpolation: 'linear',
@@ -209,16 +210,16 @@ export async function createSupporting(
   const probsBuf = new Float32Array(vocabLen);
 
   const dispose = () => {
-    oriPre.dispose();
-    dewPre.dispose();
-    tabPre.dispose();
+    orientationPreprocessor.dispose();
+    dewarpPreprocessor.dispose();
+    tablePreprocessor.dispose();
     tensors.forEach((t) => t.dispose());
     model.dispose();
   };
 
-  const detectOrientationWorklet = (input: ImageBuffer): Orientation => {
+  const detectOrientationWorklet = (page: Tensor, format: ImageFormat): Orientation => {
     'worklet';
-    const tInput = oriPre.process(input);
+    const tInput = orientationPreprocessor.processTensor(page, format);
     model.execute('orientation', [tInput], [tOri]);
     tOri.getData(oriBuf);
     const cls = argmaxRange(oriBuf, 0, oriOutLen);
@@ -232,43 +233,43 @@ export async function createSupporting(
     return { rotationCW, confidence };
   };
 
-  const dewarpWorklet = (input: ImageBuffer): ImageBuffer => {
+  // Dewarps the full-res page tensor in place: estimate the sampling field, apply
+  // it natively (cv::remap). Returns the dewarped tensor, or the input `page`
+  // unchanged when the warp is declined (caller owns whichever is returned).
+  const dewarpWorklet = (page: Tensor, format: ImageFormat): Tensor => {
     'worklet';
-    const tInput = dewPre.process(input);
+    const tInput = dewarpPreprocessor.processTensor(page, format);
     model.execute('dewarp', [tInput], [tGrid]);
-    // Apply the sampling field to the full-res page natively (cv::remap). The
-    // page-sized src/dst tensors depend on the input size, so allocate per call.
-    const ch = FORMAT_CHANNELS[input.format];
-    const tSrc = tensor('uint8', [input.height, input.width, ch]);
-    const tDst = tensor('uint8', [input.height, input.width, ch]);
+    const h = page.shape[0]!;
+    const w = page.shape[1]!;
+    const ch = FORMAT_CHANNELS[format];
+    const tDst = tensor('uint8', [h, w, ch]);
     try {
-      tSrc.setData(input.data);
-      gridSample(tSrc, tGrid, tDst);
-      const out = new Uint8Array(input.width * input.height * ch);
+      gridSample(page, tGrid, tDst);
+      const out = new Uint8Array(w * h * ch);
+      const src = new Uint8Array(w * h * ch);
       tDst.getData(out);
+      page.getData(src);
       // Degenerate-warp guard: a grid lacking page boundaries can push content
       // off-canvas, leaving a near-blank page. If the dewarp collapsed the image's
       // activity, decline it and keep the original (better an un-dewarped read than
       // zero detections).
-      if (sampledActivity(out, ch) < DEWARP_MIN_ACTIVITY_RATIO * sampledActivity(input.data, ch)) {
-        return input;
+      if (sampledActivity(out, ch) < DEWARP_MIN_ACTIVITY_RATIO * sampledActivity(src, ch)) {
+        tDst.dispose();
+        return page;
       }
-      return {
-        data: out,
-        width: input.width,
-        height: input.height,
-        format: input.format,
-        layout: input.layout,
-      };
-    } finally {
-      tSrc.dispose();
+      return tDst;
+    } catch (e) {
+      // On failure the caller can't see tDst to free it (success path returns it),
+      // so release it here before propagating.
       tDst.dispose();
+      throw e;
     }
   };
 
   const recognizeTableWorklet = (input: ImageBuffer): TableStructure => {
     'worklet';
-    const tInput = tabPre.process(input);
+    const tInput = tablePreprocessor.process(input);
     model.execute('table_encode', [tInput], [tFeatures]);
     tHidden.setData(zeroHidden);
     tOnehot.setData(zeroVocab);
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
index 3569210481..202fe71999 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/preprocessing.ts
@@ -1,7 +1,7 @@
 import { tensor, type Tensor } from '../../../core/tensor';
 import { matchShape } from '../../../core/modelSchema';
 
-import type { ImageBuffer } from '../image';
+import type { ImageBuffer, ImageFormat } from '../image';
 import {
   type ResizeMode,
   type InterpolationMethod,
@@ -56,6 +56,16 @@ export function createImagePreprocessor(
    * data.
    */
   process: (input: ImageBuffer) => Tensor;
+  /**
+   * Like {@link process}, but reads from a full-res image tensor (`[H, W, C]`,
+   * uint8) already on-device instead of an `ImageBuffer`, avoiding the raw-data
+   * copy. `format` supplies the source channel count and color conversion. The
+   * returned tensor is preprocessor-managed (do not dispose).
+   * @param src The full-res source image tensor in HWC layout.
+   * @param format The pixel format of `src` (for channels + color conversion).
+   * @returns A reference to the managed output tensor.
+   */
+  processTensor: (src: Tensor, format: ImageFormat) => Tensor;
   /**
    * Releases all allocated native resources.
    */
@@ -114,5 +124,27 @@ export function createImagePreprocessor(
     return tOutput;
   };
 
-  return { process, dispose };
+  const processTensor = (src: Tensor, format: ImageFormat): Tensor => {
+    'worklet';
+    const numChannels = FORMAT_CHANNELS[format];
+    const colorCode = FORMAT_CONVERSION[format].rgb;
+    const tResize = tensor('uint8', [targetH, targetW, numChannels]);
+    try {
+      src
+        .through(resize, tResize, {
+          mode: resizeMode,
+          interpolation: interpolation,
+          padValue: padValue,
+        })
+        .throughIf(colorCode !== null, cvtColor, tColor, colorCode!)
+        .through(toChannelsFirst, tChanFirst)
+        .through(normalize, tNorm, { alpha, beta })
+        .copyTo(tOutput);
+    } finally {
+      tResize.dispose();
+    }
+    return tOutput;
+  };
+
+  return { process, processTensor, dispose };
 }

From 9e859a25bd620cb087cc4e73c75c956e7d70ad77 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Wed, 1 Jul 2026 19:12:06 +0200
Subject: [PATCH 23/29] [RNE Rewrite] chore(ocr): explicit per-backend model
 presets, exports, cspell

---
 .cspell-wordlist.txt                          |  32 --
 .../react-native-executorch/src/constants.ts  |   4 +-
 packages/react-native-executorch/src/index.ts |   4 +-
 .../react-native-executorch/src/models.ts     | 274 +++++++++++++-----
 4 files changed, 202 insertions(+), 112 deletions(-)

diff --git a/.cspell-wordlist.txt b/.cspell-wordlist.txt
index 77307d75eb..aed3ee1e7d 100644
--- a/.cspell-wordlist.txt
+++ b/.cspell-wordlist.txt
@@ -275,35 +275,3 @@ pcre
 libkleidicv
 thresholding
 binarization
-dbnet
-svtr
-softmaxed
-softmax
-unclip
-cand
-parameterizes
-pyimagesearch
-letterbox
-CRNN
-CRAFT
-PaddleOCR
-EasyOCR
-cornerMean
-ctc
-Vatti
-softmaxing
-ppocrv
-PPOCRV
-ctcGreedyDecode
-dewarp
-vctx
-onehot
-slanet
-letterboxed
-redetect
-redetections
-eos
-doclayout
-dynint
-softmaxes
-hconcat
diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index 89e4d70548..45cf3290e2 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -1301,8 +1301,8 @@ export type BlazeFaceLandmark = (typeof BLAZEFACE_LANDMARKS)[number];
  * @category Types
  */
 export type CocoLandmark = (typeof COCO_LANDMARKS)[number];
-export { alphabets, symbols, PPOCR_SYMBOLS } from './ocrSymbols';
-export type { OCRLanguage } from './ocrSymbols';
+export { alphabets, symbols, PPOCR_SYMBOLS } from './extensions/cv/tasks/ocr/ocrSymbols';
+export type { OCRLanguage } from './extensions/cv/tasks/ocr/ocrSymbols';
 
 /**
  * PP-DocLayoutV3 region classes, in model output order (index = class id). Some
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 5fdfd6b7da..8ea8514eeb 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -7,8 +7,6 @@ export * from './hooks/useKeypointDetector';
 export * from './hooks/useObjectDetector';
 export * from './hooks/useTokenizer';
 export * from './hooks/useOCR';
-// Layout + supporting (orientation/dewarp/table) are internal to the document
-// pipeline — consumed by useDocumentOCR, not exposed as standalone hooks.
 export * from './hooks/useDocumentOCR';
 export * from './hooks/useResourceDownload';
 export * from './hooks/useModel';
@@ -26,7 +24,7 @@ export * from './extensions/cv/tasks/keypointDetection';
 export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/nlp/tasks/tokenization';
 export * from './extensions/cv/tasks/ocr';
-export * from './extensions/cv/tasks/supporting';
+export * from './extensions/cv/tasks/ocr/supporting';
 export * from './extensions/cv/tasks/documentOCR';
 
 // Core primitives — for library builders and power users
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 93dd03689c..e9a06836ec 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -5,7 +5,7 @@ import type { SemanticSegmentationModel } from './extensions/cv/tasks/semanticSe
 import type { KeypointDetectorModel } from './extensions/cv/tasks/keypointDetection';
 import type { InstanceSegmenterModel } from './extensions/cv/tasks/instanceSegmentation';
 import type { OCRModel, OCROptions } from './extensions/cv/tasks/ocr';
-import type { SupportingModel } from './extensions/cv/tasks/supporting';
+import type { SupportingModel } from './extensions/cv/tasks/ocr/supporting';
 import {
   IMAGENET_NORM,
   IMAGENET1K_LABELS,
@@ -539,72 +539,151 @@ const YOLO26_XLARGE_SEG_640_XNNPACK_FP32: InstanceSegmenterModel<'xyxy', CocoCla
 // Tokenizers
 // =============================================================================
 const ALL_MINILM_L6_V2_TOKENIZER = `${BASE_URL}-all-MiniLM-L6-v2/${VERSION_TAG}/tokenizer.json`;
+
+// =============================================================================
 // OCR
 // =============================================================================
-// EasyOCR (CRAFT + CRNN) and PaddleOCR (DBNet + SVTR). `detectorKind` selects the
-// box decoder (CRAFT heatmap grouping vs DBNet prob-map contouring) and the default
-// drop score; everything else — RGB input, recognizer normalization/padding, CTC
-// decode, confidence — is the shared baked contract (overridable per model via
-// `recognizerNorm`/`recognizerPadValue`/`decode`, which these built-ins leave at
-// the defaults). So each only declares its architecture, input-size `buckets`, and
-// `charset`. `charset` for EasyOCR is set per language.
 const EASYOCR_OPTS: OCROptions = {
-  detectorKind: 'craft', // CRAFT: text + affinity heatmaps grouped into lines
-  charset: alphabets.english, // overridden per language
+  detectorKind: 'craft',
+  charset: alphabets.english,
   buckets: { detect: [800, 1280], recognize: [64, 128, 256, 512] },
 };
 
 const PADDLE_PPOCRV6_OPTS: OCROptions = {
-  detectorKind: 'dbnet', // DBNet: binary probability map contoured into boxes
+  detectorKind: 'dbnet',
   charset: PPOCR_SYMBOLS,
   buckets: { detect: [640, 960, 1280], recognize: [160, 320, 480, 640, 1280] },
 };
-
-// OCR models live in dedicated per-architecture HF repos, pinned to `main`. Each
-// PTE is bucketed: it ships per-size `detect_<S>` / `recognize_<W>` methods and
-// the pipeline snaps each image to the closest bucket. Each repo also hosts a
-// `config.json` (detectorKind, buckets, dropScore, charsetUrl) and `charset.txt`;
-// for now those stay bundled in the family OCROptions above / constants rather
-// than being fetched.
 const OCR_REVISION = 'resolve/main';
-const EASYOCR_REPO = `${BASE_URL}-EasyOCR/${OCR_REVISION}`;
-const PPOCRV6_REPO = `${BASE_URL}-PP-OCRv6/${OCR_REVISION}`;
-const PP_DOCLAYOUT_REPO = `${BASE_URL}-PP-DocLayoutV3/${OCR_REVISION}`;
-const PADDLE_HELPERS_REPO = `${BASE_URL}-PaddleHelpers/${OCR_REVISION}`;
-
-type OCRBackend = 'xnnpack' | 'coreml' | 'vulkan';
-
-// EasyOCR hosts one subdir per language (`<lang>/EasyOCR_<lang>_<backend>.pte`);
-// the detector + pipeline are shared, only the recognizer head + charset change.
-const easyocr = (lang: string, backend: OCRBackend, charset: string): OCRModel => ({
-  modelPath: `${EASYOCR_REPO}/${lang}/EasyOCR_${lang}_${backend}.pte`,
-  ocrOpts: { ...EASYOCR_OPTS, charset },
-});
-const paddle = (backend: OCRBackend): OCRModel => ({
-  modelPath: `${PPOCRV6_REPO}/PP-OCRv6_${backend}.pte`,
-  ocrOpts: PADDLE_PPOCRV6_OPTS,
-});
 
-// EasyOCR ships 8 language recognizers (XNNPACK + CoreML + Vulkan each).
-// NOTE: per the export notes, zh_sim recognize is int8-fragile on XNNPACK/Vulkan
-// (large vocab) — prefer CoreML there.
-const easyLang = (lang: string, charset: string) => ({
-  XNNPACK: easyocr(lang, 'xnnpack', charset),
-  COREML: easyocr(lang, 'coreml', charset),
-  VULKAN: easyocr(lang, 'vulkan', charset),
-});
+// English
+const EASYOCR_ENGLISH_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
+};
+const EASYOCR_ENGLISH_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
+};
+const EASYOCR_ENGLISH_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
+};
+
+// Cyrillic
+const EASYOCR_CYRILLIC_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
+};
+const EASYOCR_CYRILLIC_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
+};
+const EASYOCR_CYRILLIC_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
+};
+
+// Latin
+const EASYOCR_LATIN_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
+};
+const EASYOCR_LATIN_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
+};
+const EASYOCR_LATIN_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
+};
+
+// Japanese
+const EASYOCR_JAPANESE_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
+};
+const EASYOCR_JAPANESE_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
+};
+const EASYOCR_JAPANESE_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
+};
+
+// Simplified Chinese
+const EASYOCR_ZH_SIM_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
+};
+const EASYOCR_ZH_SIM_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
+};
+const EASYOCR_ZH_SIM_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
+};
 
-const PADDLE_PPOCRV6_XNNPACK = paddle('xnnpack');
-const PADDLE_PPOCRV6_VULKAN = paddle('vulkan');
-const PADDLE_PPOCRV6_COREML = paddle('coreml');
+// Korean
+const EASYOCR_KOREAN_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
+};
+const EASYOCR_KOREAN_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
+};
+const EASYOCR_KOREAN_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
+};
+
+// Telugu
+const EASYOCR_TELUGU_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
+};
+const EASYOCR_TELUGU_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
+};
+const EASYOCR_TELUGU_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
+};
+
+// Kannada
+const EASYOCR_KANNADA_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_xnnpack.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
+};
+const EASYOCR_KANNADA_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_coreml.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
+};
+const EASYOCR_KANNADA_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_vulkan.pte`,
+  ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
+};
+
+const PADDLE_PPOCRV6_XNNPACK: OCRModel = {
+  modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_xnnpack.pte`,
+  ocrOpts: PADDLE_PPOCRV6_OPTS,
+};
+const PADDLE_PPOCRV6_COREML: OCRModel = {
+  modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_coreml.pte`,
+  ocrOpts: PADDLE_PPOCRV6_OPTS,
+};
+const PADDLE_PPOCRV6_VULKAN: OCRModel = {
+  modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_vulkan.pte`,
+  ocrOpts: PADDLE_PPOCRV6_OPTS,
+};
 
 // =============================================================================
-// Document layout — PP-DocLayoutV3 (RT-DETR, 25 region classes, fixed 800x800).
+// Document layout — PP-DocLayoutV3
 // =============================================================================
-// PP-DocLayoutV3 speaks the objectDetection contract (the RT-DETR decode is baked
-// in-graph): RGB 800x800, x/255 with NO mean/std, stretch. NMS is disabled
-// (iouThreshold 1) because RT-DETR is set-prediction and standard NMS would merge
-// legitimately nested regions (e.g. a table inside a figure).
 const PP_DOCLAYOUT_OPTS = {
   labels: DOC_LAYOUT_LABELS,
   boxFormat: 'xyxy' as const,
@@ -615,27 +694,40 @@ const PP_DOCLAYOUT_OPTS = {
   defaultConfidenceThreshold: 0.3,
   defaultIouThreshold: 1.0,
 };
-const layout = (backend: OCRBackend): ObjectDetectorModel<'xyxy', DocLayoutLabel> => ({
-  modelPath: `${PP_DOCLAYOUT_REPO}/PP-DocLayoutV3_${backend}.pte`,
+const PP_DOCLAYOUT_XNNPACK: ObjectDetectorModel<'xyxy', DocLayoutLabel> = {
+  modelPath: `${BASE_URL}-pp-doclayout-v3/${OCR_REVISION}/PP-DocLayoutV3_xnnpack.pte`,
+  opts: PP_DOCLAYOUT_OPTS,
+};
+const PP_DOCLAYOUT_COREML: ObjectDetectorModel<'xyxy', DocLayoutLabel> = {
+  modelPath: `${BASE_URL}-pp-doclayout-v3/${OCR_REVISION}/PP-DocLayoutV3_coreml.pte`,
+  opts: PP_DOCLAYOUT_OPTS,
+};
+const PP_DOCLAYOUT_VULKAN: ObjectDetectorModel<'xyxy', DocLayoutLabel> = {
+  modelPath: `${BASE_URL}-pp-doclayout-v3/${OCR_REVISION}/PP-DocLayoutV3_vulkan.pte`,
   opts: PP_DOCLAYOUT_OPTS,
-});
-const PP_DOCLAYOUT_XNNPACK = layout('xnnpack');
-const PP_DOCLAYOUT_VULKAN = layout('vulkan');
-const PP_DOCLAYOUT_COREML = layout('coreml');
+};
 
 // =============================================================================
-// Table structure — SLANet_plus (in the fused supporting PTE: table_encode +
-// table_decode_step). Also carries orientation/dewarp methods (same file).
+// Document helpers - Paddle supporting models
 // =============================================================================
-const supporting = (backend: OCRBackend): SupportingModel => ({
-  modelPath: `${PADDLE_HELPERS_REPO}/PaddleHelpers_${backend}.pte`,
+const PP_TABLE_XNNPACK: SupportingModel = {
+  modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_xnnpack.pte`,
+  vocab: SLANET_STRUCTURE_VOCAB,
+  eosTokenId: 49,
+  maxSteps: 501,
+};
+const PP_TABLE_COREML: SupportingModel = {
+  modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_coreml.pte`,
   vocab: SLANET_STRUCTURE_VOCAB,
-  eosTokenId: 49, // SLANet '</s>' token id
-  maxSteps: 501, // SLANet max_text_length (500) + 1
-});
-const PP_TABLE_XNNPACK = supporting('xnnpack');
-const PP_TABLE_VULKAN = supporting('vulkan');
-const PP_TABLE_COREML = supporting('coreml');
+  eosTokenId: 49,
+  maxSteps: 501,
+};
+const PP_TABLE_VULKAN: SupportingModel = {
+  modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_vulkan.pte`,
+  vocab: SLANET_STRUCTURE_VOCAB,
+  eosTokenId: 49,
+  maxSteps: 501,
+};
 
 /**
  * Registry of pre-configured ExecuTorch models.
@@ -843,14 +935,46 @@ export const models = {
   },
   ocr: {
     EASYOCR: {
-      ENGLISH: easyLang('english', alphabets.english),
-      CYRILLIC: easyLang('cyrillic', alphabets.cyrillic),
-      LATIN: easyLang('latin', alphabets.latin),
-      JAPANESE: easyLang('japanese', alphabets.japanese),
-      ZH_SIM: easyLang('zh_sim', alphabets.zh_sim),
-      KOREAN: easyLang('korean', alphabets.korean),
-      TELUGU: easyLang('telugu', alphabets.telugu),
-      KANNADA: easyLang('kannada', alphabets.kannada),
+      ENGLISH: {
+        XNNPACK: EASYOCR_ENGLISH_XNNPACK,
+        COREML: EASYOCR_ENGLISH_COREML,
+        VULKAN: EASYOCR_ENGLISH_VULKAN,
+      },
+      CYRILLIC: {
+        XNNPACK: EASYOCR_CYRILLIC_XNNPACK,
+        COREML: EASYOCR_CYRILLIC_COREML,
+        VULKAN: EASYOCR_CYRILLIC_VULKAN,
+      },
+      LATIN: {
+        XNNPACK: EASYOCR_LATIN_XNNPACK,
+        COREML: EASYOCR_LATIN_COREML,
+        VULKAN: EASYOCR_LATIN_VULKAN,
+      },
+      JAPANESE: {
+        XNNPACK: EASYOCR_JAPANESE_XNNPACK,
+        COREML: EASYOCR_JAPANESE_COREML,
+        VULKAN: EASYOCR_JAPANESE_VULKAN,
+      },
+      ZH_SIM: {
+        XNNPACK: EASYOCR_ZH_SIM_XNNPACK,
+        COREML: EASYOCR_ZH_SIM_COREML,
+        VULKAN: EASYOCR_ZH_SIM_VULKAN,
+      },
+      KOREAN: {
+        XNNPACK: EASYOCR_KOREAN_XNNPACK,
+        COREML: EASYOCR_KOREAN_COREML,
+        VULKAN: EASYOCR_KOREAN_VULKAN,
+      },
+      TELUGU: {
+        XNNPACK: EASYOCR_TELUGU_XNNPACK,
+        COREML: EASYOCR_TELUGU_COREML,
+        VULKAN: EASYOCR_TELUGU_VULKAN,
+      },
+      KANNADA: {
+        XNNPACK: EASYOCR_KANNADA_XNNPACK,
+        COREML: EASYOCR_KANNADA_COREML,
+        VULKAN: EASYOCR_KANNADA_VULKAN,
+      },
     },
     PADDLE: {
       PPOCRV6_SMALL: {

From b31d9241fe977e1cb83db68e3f2020632df572f7 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 00:10:08 +0200
Subject: [PATCH 24/29] [RNE Rewrite] refactor(cv): split OCR native ops, add
 warpQuad/warpByGrid

---
 .../cpp/extensions/cv/image_ops.cpp           | 251 ++++++++++++++
 .../cpp/extensions/cv/image_ops.h             |   5 +
 .../cpp/extensions/cv/install.cpp             |   9 +-
 .../cpp/extensions/cv/ocr_ops.cpp             | 309 ++++++++----------
 .../cpp/extensions/cv/ocr_ops.h               |  15 +-
 .../cpp/extensions/cv/text_boxes_ops.cpp      | 266 ---------------
 .../cpp/extensions/cv/text_boxes_ops.h        |   8 -
 7 files changed, 410 insertions(+), 453 deletions(-)
 delete mode 100644 packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
 delete mode 100644 packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h

diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index c0e5b513f9..eda54f2d15 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -811,4 +811,255 @@ void install_rotate(jsi::Runtime &rt, jsi::Object &module) {
     module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
 }
 
+// ------------------------------- warpByGrid --------------------------------
+// Warp `src` through a backward sampling field (torch grid_sample step of a
+// geometric dewarp) into `dst` via cv::remap. grid is [..,2,gH,gW], normalized
+// to [-1,1] with align_corners=true (channel 0 = x, 1 = y).
+void install_warpByGrid(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "warpByGrid";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
+        if (count != 3) {
+            throw jsi::JSError(rt, "Usage: warpByGrid(src, grid, dst)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt) ||
+            !args[2].isObject() || !args[2].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpByGrid: src, grid, and dst must be Tensors");
+        }
+
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto grid = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[2].asObject(rt).getHostObject<TensorHostObject>(rt);
+
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "warpByGrid: In-place operations (src == dst) are not supported.");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "warpByGrid: src and dst must be uint8");
+        }
+        if (grid->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "warpByGrid: grid must be float32");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "warpByGrid: src and dst must be [H, W, C]");
+        }
+        if (src->shape_ != dst->shape_) {
+            throw jsi::JSError(rt, "warpByGrid: src and dst must have the same shape");
+        }
+        // grid is the torch grid_sample field [..,2,gH,gW], channel 0 = x, 1 = y,
+        // normalized to [-1,1] with align_corners=true.
+        const auto &gs = grid->shape_;
+        if (gs.size() < 3 || gs[gs.size() - 3] != 2) {
+            throw jsi::JSError(rt, "warpByGrid: grid must be [..,2,gH,gW]");
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        std::shared_lock<std::shared_mutex> gridLock(grid->mutex_, std::try_to_lock);
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock() || !gridLock.owns_lock() || !dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpByGrid: a tensor is currently in use");
+        }
+        if (!src->data_ || !grid->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "warpByGrid: a tensor has been disposed");
+        }
+
+        const int32_t h = src->shape_[0];
+        const int32_t w = src->shape_[1];
+        const int32_t channels = src->shape_[2];
+        const int32_t gridH = gs[gs.size() - 2];
+        const int32_t gridW = gs[gs.size() - 1];
+        const int32_t plane = gridH * gridW;
+        const auto *g = reinterpret_cast<const float *>(grid->data_.get());
+
+        // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
+        auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
+            const int32_t x0 = std::clamp(static_cast<int32_t>(std::floor(gx)), 0, gridW - 1);
+            const int32_t y0 = std::clamp(static_cast<int32_t>(std::floor(gy)), 0, gridH - 1);
+            const int32_t x1 = std::min(x0 + 1, gridW - 1);
+            const int32_t y1 = std::min(y0 + 1, gridH - 1);
+            const float dx = gx - static_cast<float>(x0);
+            const float dy = gy - static_cast<float>(y0);
+            const int32_t base = c * plane;
+            const float top = g[base + y0 * gridW + x0] +
+                              (g[base + y0 * gridW + x1] - g[base + y0 * gridW + x0]) * dx;
+            const float bot = g[base + y1 * gridW + x0] +
+                              (g[base + y1 * gridW + x1] - g[base + y1 * gridW + x0]) * dx;
+            return top + (bot - top) * dy;
+        };
+
+        ::cv::Mat mapX(h, w, CV_32F);
+        ::cv::Mat mapY(h, w, CV_32F);
+        for (int32_t oy = 0; oy < h; ++oy) {
+            const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
+                                         static_cast<float>(gridH - 1)
+                                   : 0.0f;
+            auto *rowX = mapX.ptr<float>(oy);
+            auto *rowY = mapY.ptr<float>(oy);
+            for (int32_t ox = 0; ox < w; ++ox) {
+                const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
+                                             static_cast<float>(gridW - 1)
+                                       : 0.0f;
+                const float nx = sampleGrid(0, gx, gy); // [-1,1]
+                const float ny = sampleGrid(1, gx, gy);
+                rowX[ox] = ((nx + 1.0f) / 2.0f) * static_cast<float>(w - 1);
+                rowY[ox] = ((ny + 1.0f) / 2.0f) * static_cast<float>(h - 1);
+            }
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(h, w, cvType, src->data_.get());
+        ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
+        try {
+            ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("warpByGrid: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[2]);
+    };
+    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
+}
+
+// ------------------------------- warpQuad ----------------------------------
+// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
+// resize-to-height + pad/align). Used by the OCR recognizer to normalize a
+// detected text box into the fixed recognizer canvas.
+void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "warpQuad";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 4) {
+            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
+        }
+        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
+        }
+        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
+            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
+        }
+        if (!args[3].isObject()) {
+            throw jsi::JSError(rt, "warpQuad: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
+        if (src.get() == dst.get()) {
+            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
+        }
+        auto quadArr = args[2].asObject(rt).asArray(rt);
+        auto opts = args[3].asObject(rt);
+
+        if (quadArr.length(rt) != 8) {
+            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
+        }
+        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
+        }
+        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
+            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
+        }
+        if (src->shape_[2] != dst->shape_[2]) {
+            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
+        }
+
+        const int32_t channels = src->shape_[2];
+        const int32_t recH = dst->shape_[0];
+        const int32_t bucketW = dst->shape_[1];
+
+        if (!opts.hasProperty(rt, "contentWidth") ||
+            !opts.getProperty(rt, "contentWidth").isNumber()) {
+            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
+        }
+        const int32_t contentWidth =
+            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
+                       bucketW);
+        const std::string padMode = opts.getProperty(rt, "padMode").asString(rt).utf8(rt);
+        const double padValue = opts.getProperty(rt, "padValue").asNumber();
+        const std::string align = opts.getProperty(rt, "align").asString(rt).utf8(rt);
+        // offsetX >= 0 places content at that x (overriding align); clear=false skips
+        // wiping dst first, so successive warps compose into one canvas (glyph strips).
+        const auto offsetXOpt = static_cast<int32_t>(opts.getProperty(rt, "offsetX").asNumber());
+        const bool clear = opts.getProperty(rt, "clear").asBool();
+
+        std::array<::cv::Point2f, 4> quad;
+        for (std::size_t i = 0; i < 8; ++i) {
+            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
+                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
+            }
+        }
+        for (std::size_t i = 0; i < 4; ++i) {
+            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
+                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
+        }
+        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
+        if (!dstLock.owns_lock()) {
+            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
+        }
+        if (!src->data_ || !dst->data_) {
+            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
+        }
+
+        const int cvType = CV_MAKETYPE(CV_8U, channels);
+        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
+        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
+
+        try {
+            const std::array<::cv::Point2f, 4> dstPts = {
+                ::cv::Point2f{0.0f, 0.0f},
+                {static_cast<float>(contentWidth), 0.0f},
+                {static_cast<float>(contentWidth), static_cast<float>(recH)},
+                {0.0f, static_cast<float>(recH)}};
+            const std::array<::cv::Point2f, 4> srcPts = {quad[0], quad[1], quad[2], quad[3]};
+            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts.data(), dstPts.data());
+            ::cv::Mat content;
+            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
+                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
+
+            ::cv::Scalar padColor;
+            if (padMode == "cornerMean") {
+                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
+                ::cv::Scalar acc(0, 0, 0, 0);
+                const std::array<::cv::Rect, 4> rects = {
+                    ::cv::Rect(0, 0, patch, patch),
+                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
+                    ::cv::Rect(0, recH - patch, patch, patch),
+                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
+                for (const auto &r : rects) {
+                    acc += ::cv::mean(content(r));
+                }
+                padColor = acc / 4.0;
+            } else {
+                padColor = ::cv::Scalar::all(padValue);
+            }
+
+            if (clear) {
+                dstMat.setTo(padColor);
+            }
+            int32_t offsetX = offsetXOpt;
+            if (offsetX < 0) {
+                offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
+            }
+            if (offsetX < bucketW) {
+                const int32_t copyW = std::min(contentWidth, bucketW - offsetX);
+                content(::cv::Rect(0, 0, copyW, recH))
+                    .copyTo(dstMat(::cv::Rect(offsetX, 0, copyW, recH)));
+            }
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
+        }
+        return jsi::Value(rt, args[1]);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             4, fnBody));
+}
+
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
index 8c91c16878..a694809aac 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
@@ -10,4 +10,9 @@ void install_toChannelsLast(facebook::jsi::Runtime &rt, facebook::jsi::Object &m
 void install_normalize(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_applyColormap(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_rotate(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+// Warp `src` through a low-res [-1,1] sampling grid into `dst` (dewarp / remap).
+void install_warpByGrid(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
+// resize-to-height + pad/align).
+void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/install.cpp b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
index a228a55654..ba32a96041 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/install.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/install.cpp
@@ -2,7 +2,6 @@
 #include "box_ops.h"
 #include "image_ops.h"
 #include "ocr_ops.h"
-#include "text_boxes_ops.h"
 
 namespace rnexecutorch::extensions::cv {
 namespace jsi = facebook::jsi;
@@ -17,16 +16,16 @@ void install(facebook::jsi::Runtime &rt, facebook::jsi::Object &module) {
     image_ops::install_normalize(rt, cvModule);
     image_ops::install_applyColormap(rt, cvModule);
     image_ops::install_rotate(rt, cvModule);
+    image_ops::install_warpByGrid(rt, cvModule);
+    image_ops::install_warpQuad(rt, cvModule);
 
     box_ops::install_nms(rt, cvModule);
     box_ops::install_restrictToBox(rt, cvModule);
 
-    ocr_ops::install_extractTextBoxes(rt, cvModule);
+    ocr_ops::install_extractCraftTextBoxes(rt, cvModule);
+    ocr_ops::install_extractDbnetTextBoxes(rt, cvModule);
     ocr_ops::install_ctcGreedyDecode(rt, cvModule);
 
-    text_boxes_ops::install_gridSample(rt, cvModule);
-    text_boxes_ops::install_warpQuad(rt, cvModule);
-
     module.setProperty(rt, "cv", cvModule);
 }
 } // namespace rnexecutorch::extensions::cv
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index 68f13c545d..f11c2c2f83 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -10,7 +10,6 @@
 #include <numeric>
 #include <opencv2/core/check.hpp>
 #include <optional>
-#include <span>
 #include <stdexcept>
 #include <tuple>
 #include <unordered_set>
@@ -67,22 +66,6 @@ ::cv::Point2f rotateAround(const ::cv::Point2f &p, const ::cv::Point2f &ctr, flo
 }
 
 // ------------------------------ CRAFT branch -------------------------------
-std::pair<::cv::Mat, ::cv::Mat> interleavedToMats(std::span<const float> data, ::cv::Size size) {
-    ::cv::Mat textMap(size, CV_32F);
-    ::cv::Mat affinityMap(size, CV_32F);
-    const auto w = static_cast<std::size_t>(size.width);
-    for (std::size_t i = 0; i < data.size(); ++i) {
-        const auto x = static_cast<int32_t>((i / 2) % w);
-        const auto y = static_cast<int32_t>((i / 2) / w);
-        if (i % 2 == 0) {
-            textMap.at<float>(y, x) = data[i];
-        } else {
-            affinityMap.at<float>(y, x) = data[i];
-        }
-    }
-    return {textMap, affinityMap};
-}
-
 void dilateComponent(::cv::Mat &segMap, const ::cv::Mat &stats, int32_t i, int32_t area,
                      int32_t imgW, int32_t imgH) {
     const int32_t x = stats.at<int32_t>(i, ::cv::CC_STAT_LEFT);
@@ -135,9 +118,17 @@ std::optional<Box> boxFromComponent(const ::cv::Mat &textMap, const ::cv::Mat &l
     return box;
 }
 
-std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityMap,
-                                        float textThreshold, float linkThreshold,
-                                        float lowTextThreshold) {
+// CRAFT text+affinity maps -> component boxes, in two modes:
+//   - line grouping (charLevel=false): affinity is ADDED to the text map so
+//     adjacent glyphs link into one region; boxes keep their rotated-rect angle.
+//   - char level (charLevel=true): affinity is SUBTRACTED to BREAK those links,
+//     and the mask is eroded/dilated to clean up, yielding one upright box per
+//     glyph (used by the per-column pass that reads stacked text glyph by glyph;
+//     mirrors the old VerticalDetector's single-character path).
+// Everything after the combine step (binarize -> connected components -> one box
+// per component) is shared. charLevel boxes are forced upright (angle 0).
+std::vector<Box> componentBoxes(::cv::Mat &textMap, ::cv::Mat &affinityMap, float textThreshold,
+                                float linkThreshold, float lowTextThreshold, bool charLevel) {
     const int32_t imgH = textMap.rows;
     const int32_t imgW = textMap.cols;
     ::cv::Mat textScore;
@@ -145,11 +136,22 @@ std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityM
     ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
     ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
                     ::cv::THRESH_BINARY);
-    ::cv::Mat comb = textScore + affinityScore;
-    ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_BINARY);
+
+    ::cv::Mat comb;
+    if (charLevel) {
+        comb = textScore - affinityScore; // subtract to separate adjacent glyphs
+        ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_TOZERO);
+        ::cv::threshold(comb, comb, 1.0, 1.0, ::cv::THRESH_TRUNC);
+        ::cv::Mat kernel = ::cv::getStructuringElement(::cv::MORPH_RECT, ::cv::Size(3, 3));
+        ::cv::erode(comb, comb, kernel, ::cv::Point(-1, -1), 1);
+        ::cv::dilate(comb, comb, kernel, ::cv::Point(-1, -1), 4);
+    } else {
+        comb = textScore + affinityScore; // add to link adjacent glyphs into lines
+        ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_BINARY);
+    }
+
     ::cv::Mat binary;
     comb.convertTo(binary, CV_8UC1);
-
     ::cv::Mat labels;
     ::cv::Mat stats;
     ::cv::Mat centroids;
@@ -160,6 +162,9 @@ std::vector<Box> getDetBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityM
     for (int32_t i = 1; i < nLabels; ++i) {
         auto box = boxFromComponent(textMap, labels, stats, i, imgW, imgH, lowTextThreshold);
         if (box) {
+            if (charLevel) {
+                box->angle = 0.0f; // glyphs are read upright, never rotated
+            }
             boxes.push_back(*box);
         }
     }
@@ -331,61 +336,19 @@ std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
     return filtered;
 }
 
-// Char-level CRAFT extraction: one upright box per glyph, no line grouping. The
-// affinity map is SUBTRACTED from the text map to break the links between
-// adjacent characters (the opposite of the grouped path, which adds them), then
-// the components are eroded/dilated to clean up before labelling. Used by the
-// second, per-column detection pass that reads upright stacked text glyph by
-// glyph. Mirrors the old VerticalDetector's single-character path.
-std::vector<Box> getCharBoxesFromTextMap(::cv::Mat &textMap, ::cv::Mat &affinityMap,
-                                         float textThreshold, float linkThreshold,
-                                         float lowTextThreshold) {
-    const int32_t imgH = textMap.rows;
-    const int32_t imgW = textMap.cols;
-    ::cv::Mat textScore;
-    ::cv::Mat affinityScore;
-    ::cv::threshold(textMap, textScore, static_cast<double>(textThreshold), 1.0, ::cv::THRESH_BINARY);
-    ::cv::threshold(affinityMap, affinityScore, static_cast<double>(linkThreshold), 1.0,
-                    ::cv::THRESH_BINARY);
-    ::cv::Mat comb = textScore - affinityScore; // subtract to separate adjacent glyphs
-    ::cv::threshold(comb, comb, 0.0, 1.0, ::cv::THRESH_TOZERO);
-    ::cv::threshold(comb, comb, 1.0, 1.0, ::cv::THRESH_TRUNC);
-    ::cv::Mat kernel = ::cv::getStructuringElement(::cv::MORPH_RECT, ::cv::Size(3, 3));
-    ::cv::erode(comb, comb, kernel, ::cv::Point(-1, -1), 1);
-    ::cv::dilate(comb, comb, kernel, ::cv::Point(-1, -1), 4);
-
-    ::cv::Mat binary;
-    comb.convertTo(binary, CV_8UC1);
-    ::cv::Mat labels;
-    ::cv::Mat stats;
-    ::cv::Mat centroids;
-    const int32_t nLabels = ::cv::connectedComponentsWithStats(binary, labels, stats, centroids, 4);
-
-    std::vector<Box> boxes;
-    boxes.reserve(static_cast<std::size_t>(nLabels));
-    for (int32_t i = 1; i < nLabels; ++i) {
-        auto box = boxFromComponent(textMap, labels, stats, i, imgW, imgH, lowTextThreshold);
-        if (box) {
-            box->angle = 0.0f; // glyphs are read upright, never rotated
-            boxes.push_back(*box);
-        }
-    }
-    return boxes;
-}
-
 // CRAFT half-res heatmap (text+affinity interleaved) -> oriented quads in
 // detector-input pixels; restoreRatio scales the half-res boxes back up. With
 // charLevel the boxes are individual upright glyphs (no grouping); otherwise
-// they are grouped reading-ordered lines.
-std::vector<Quad> extractCraft(std::span<const float> data, int32_t heatW, int32_t heatH,
-                               float textThreshold, float linkThreshold, float lowTextThreshold,
-                               float restoreRatio, bool charLevel) {
-    auto [textMap, affinityMap] = interleavedToMats(data, ::cv::Size(heatW, heatH));
-    std::vector<Box> boxes =
-        charLevel ? getCharBoxesFromTextMap(textMap, affinityMap, textThreshold, linkThreshold,
-                                            lowTextThreshold)
-                  : getDetBoxesFromTextMap(textMap, affinityMap, textThreshold, linkThreshold,
-                                           lowTextThreshold);
+// they are grouped reading-ordered lines. `data` points at heatW*heatH*2 floats.
+std::vector<Quad> extractCraft(float *data, int32_t heatW, int32_t heatH, float textThreshold,
+                               float linkThreshold, float lowTextThreshold, float restoreRatio,
+                               bool charLevel) {
+    // Deinterleave the [text, affinity] channels of the half-res heatmap.
+    ::cv::Mat interleaved(heatH, heatW, CV_32FC2, data);
+    std::array<::cv::Mat, 2> channels;
+    ::cv::split(interleaved, channels);
+    std::vector<Box> boxes = componentBoxes(channels[0], channels[1], textThreshold, linkThreshold,
+                                            lowTextThreshold, charLevel);
     for (auto &b : boxes) {
         b.x0 *= restoreRatio;
         b.y0 *= restoreRatio;
@@ -424,22 +387,12 @@ std::vector<Quad> extractCraft(std::span<const float> data, int32_t heatW, int32
 }
 
 // ------------------------------ DBNet branch -------------------------------
-// DBNet prob map [H,W] -> oriented quads.
-std::vector<Quad> extractDbnet(const ::cv::Mat &probIn, float binThreshold, float boxThreshold,
-                               float unclipRatio, int32_t minBoxSide, int32_t maxCandidates,
-                               bool applySigmoid) {
-    const int32_t w = probIn.cols;
-    const int32_t h = probIn.rows;
-    // The caller declares (from the model's export contract) whether the head
-    // emits raw logits (apply sigmoid) or already-normalized probabilities.
-    ::cv::Mat prob;
-    if (applySigmoid) {
-        ::cv::Mat neg;
-        ::cv::exp(-probIn, neg);
-        prob = 1.0 / (1.0 + neg);
-    } else {
-        prob = probIn;
-    }
+// DBNet prob map [H,W] -> oriented quads. The map must be post-sigmoid
+// probabilities — any activation is baked into the model's export.
+std::vector<Quad> extractDbnet(const ::cv::Mat &prob, float binThreshold, float boxThreshold,
+                               float unclipRatio, int32_t minBoxSide, int32_t maxCandidates) {
+    const int32_t w = prob.cols;
+    const int32_t h = prob.rows;
 
     ::cv::Mat bitmap;
     ::cv::threshold(prob, bitmap, static_cast<double>(binThreshold), 255, ::cv::THRESH_BINARY);
@@ -535,82 +488,111 @@ jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
 
 } // namespace
 
-void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
-    const auto *name = "extractTextBoxes";
+void install_extractCraftTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "extractCraftTextBoxes";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
                      size_t count) -> jsi::Value {
         if (count != 2) {
-            throw jsi::JSError(rt, "Usage: extractTextBoxes(src, options)");
+            throw jsi::JSError(rt, "Usage: extractCraftTextBoxes(src, options)");
         }
         if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "extractTextBoxes: src must be a Tensor");
+            throw jsi::JSError(rt, "extractCraftTextBoxes: src must be a Tensor");
         }
         if (!args[1].isObject()) {
-            throw jsi::JSError(rt, "extractTextBoxes: options must be an object");
+            throw jsi::JSError(rt, "extractCraftTextBoxes: options must be an object");
         }
         auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
         auto opts = args[1].asObject(rt);
-
         if (src->dtype_ != rnexecutorch::core::types::DType::float32) {
-            throw jsi::JSError(rt, "extractTextBoxes: src must be a float32 Tensor");
+            throw jsi::JSError(rt, "extractCraftTextBoxes: src must be a float32 Tensor");
         }
-        if (!opts.hasProperty(rt, "mode") || !opts.getProperty(rt, "mode").isString()) {
-            throw jsi::JSError(rt, "extractTextBoxes: options.mode is required and must be a string");
-        }
-        const std::string mode = opts.getProperty(rt, "mode").asString(rt).utf8(rt);
 
         std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
         if (!srcLock.owns_lock()) {
-            throw jsi::JSError(rt, "extractTextBoxes: src tensor is currently in use");
+            throw jsi::JSError(rt, "extractCraftTextBoxes: src tensor is currently in use");
         }
         if (!src->data_) {
-            throw jsi::JSError(rt, "extractTextBoxes: src tensor has been disposed");
+            throw jsi::JSError(rt, "extractCraftTextBoxes: src tensor has been disposed");
+        }
+        auto *dataPtr = reinterpret_cast<float *>(src->data_.get());
+
+        // src is [1,Hd,Wd,2] or [Hd,Wd,2] interleaved (text, affinity), half-res.
+        const auto &s = src->shape_;
+        if (s.size() < 3 || s.back() != 2) {
+            throw jsi::JSError(rt, "extractCraftTextBoxes: src must be [..,Hd,Wd,2]");
         }
+        const int32_t heatW = s[s.size() - 2];
+        const int32_t heatH = s[s.size() - 3];
+        const double targetH = opts.getProperty(rt, "targetHeight").asNumber();
+        const float restoreRatio = static_cast<float>(targetH) / static_cast<float>(heatH);
+        // Required option — default values live in the TypeScript wrapper layer.
+        const bool charLevel = opts.getProperty(rt, "charLevel").asBool();
 
+        std::vector<Quad> quads;
+        try {
+            quads = extractCraft(
+                dataPtr, heatW, heatH,
+                static_cast<float>(opts.getProperty(rt, "textThreshold").asNumber()),
+                static_cast<float>(opts.getProperty(rt, "linkThreshold").asNumber()),
+                static_cast<float>(opts.getProperty(rt, "lowTextThreshold").asNumber()),
+                restoreRatio, charLevel);
+        } catch (const ::cv::Exception &e) {
+            throw jsi::JSError(rt, std::string("extractCraftTextBoxes: OpenCV error: ") + e.what());
+        }
+        return quadsToArray(rt, quads);
+    };
+    module.setProperty(rt, name,
+                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
+                                                             2, fnBody));
+}
+
+void install_extractDbnetTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
+    const auto *name = "extractDbnetTextBoxes";
+    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
+                     size_t count) -> jsi::Value {
+        if (count != 2) {
+            throw jsi::JSError(rt, "Usage: extractDbnetTextBoxes(src, options)");
+        }
+        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: src must be a Tensor");
+        }
+        if (!args[1].isObject()) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: options must be an object");
+        }
+        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
+        auto opts = args[1].asObject(rt);
+        if (src->dtype_ != rnexecutorch::core::types::DType::float32) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: src must be a float32 Tensor");
+        }
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: src tensor is currently in use");
+        }
+        if (!src->data_) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: src tensor has been disposed");
+        }
         auto *dataPtr = reinterpret_cast<float *>(src->data_.get());
-        std::span<const float> data(dataPtr, src->numel_);
+
+        // src is [1,1,H,W] or [H,W] probability map (full-res).
+        const auto &s = src->shape_;
+        if (s.size() < 2) {
+            throw jsi::JSError(rt, "extractDbnetTextBoxes: src must be [..,H,W]");
+        }
+        const int32_t w = s[s.size() - 1];
+        const int32_t h = s[s.size() - 2];
 
         std::vector<Quad> quads;
         try {
-            if (mode == "craft") {
-                // src is [1,Hd,Wd,2] or [Hd,Wd,2] interleaved (text, affinity), half-res.
-                const auto &s = src->shape_;
-                if (s.size() < 3 || s.back() != 2) {
-                    throw jsi::JSError(rt, "extractTextBoxes: craft src must be [..,Hd,Wd,2]");
-                }
-                const int32_t heatW = s[s.size() - 2];
-                const int32_t heatH = s[s.size() - 3];
-                const double targetH = opts.getProperty(rt, "targetHeight").asNumber();
-                const float restoreRatio = static_cast<float>(targetH) / static_cast<float>(heatH);
-                const bool charLevel =
-                    opts.hasProperty(rt, "charLevel") && opts.getProperty(rt, "charLevel").asBool();
-                quads = extractCraft(
-                    data, heatW, heatH,
-                    static_cast<float>(opts.getProperty(rt, "textThreshold").asNumber()),
-                    static_cast<float>(opts.getProperty(rt, "linkThreshold").asNumber()),
-                    static_cast<float>(opts.getProperty(rt, "lowTextThreshold").asNumber()), restoreRatio,
-                    charLevel);
-            } else if (mode == "dbnet") {
-                // src is [1,1,H,W] or [H,W] probability map (full-res).
-                const auto &s = src->shape_;
-                if (s.size() < 2) {
-                    throw jsi::JSError(rt, "extractTextBoxes: dbnet src must be [..,H,W]");
-                }
-                const int32_t w = s[s.size() - 1];
-                const int32_t h = s[s.size() - 2];
-                ::cv::Mat prob(h, w, CV_32F, dataPtr);
-                quads = extractDbnet(
-                    prob, static_cast<float>(opts.getProperty(rt, "binThreshold").asNumber()),
-                    static_cast<float>(opts.getProperty(rt, "boxThreshold").asNumber()),
-                    static_cast<float>(opts.getProperty(rt, "unclipRatio").asNumber()),
-                    static_cast<int32_t>(opts.getProperty(rt, "minBoxSide").asNumber()),
-                    static_cast<int32_t>(opts.getProperty(rt, "maxCandidates").asNumber()),
-                    opts.getProperty(rt, "applySigmoid").asBool());
-            } else {
-                throw jsi::JSError(rt, "extractTextBoxes: unknown mode '" + mode + "'");
-            }
+            ::cv::Mat prob(h, w, CV_32F, dataPtr);
+            quads = extractDbnet(
+                prob, static_cast<float>(opts.getProperty(rt, "binThreshold").asNumber()),
+                static_cast<float>(opts.getProperty(rt, "boxThreshold").asNumber()),
+                static_cast<float>(opts.getProperty(rt, "unclipRatio").asNumber()),
+                static_cast<int32_t>(opts.getProperty(rt, "minBoxSide").asNumber()),
+                static_cast<int32_t>(opts.getProperty(rt, "maxCandidates").asNumber()));
         } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("extractTextBoxes: OpenCV error: ") + e.what());
+            throw jsi::JSError(rt, std::string("extractDbnetTextBoxes: OpenCV error: ") + e.what());
         }
         return quadsToArray(rt, quads);
     };
@@ -620,21 +602,28 @@ void install_extractTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
 }
 
 // --------------------------- ctcGreedyDecode -------------------------------
+// Per-timestep argmax + max value over [..,T,V] logits. `values` are the raw
+// max activations; if a caller needs probabilities it softmaxes the tensor (via
+// the math.softmax op) before decoding — this op takes no options.
 void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
     const auto *name = "ctcGreedyDecode";
     auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
                      size_t count) -> jsi::Value {
-        if (count != 2) {
-            throw jsi::JSError(rt, "Usage: ctcGreedyDecode(src, options)");
+        if (count != 1) {
+            throw jsi::JSError(rt, "Usage: ctcGreedyDecode(src)");
         }
         if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
             throw jsi::JSError(rt, "ctcGreedyDecode: src must be a Tensor");
         }
-        if (!args[1].isObject()) {
-            throw jsi::JSError(rt, "ctcGreedyDecode: options must be an object");
-        }
         auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto opts = args[1].asObject(rt);
+
+        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
+        if (!srcLock.owns_lock()) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor is currently in use");
+        }
+        if (!src->data_) {
+            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor has been disposed");
+        }
 
         if (src->dtype_ != rnexecutorch::core::types::DType::float32) {
             throw jsi::JSError(rt, "ctcGreedyDecode: src must be a float32 Tensor");
@@ -651,15 +640,6 @@ void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
             throw jsi::JSError(rt, "ctcGreedyDecode: numel must be a multiple of the vocab dim");
         }
         const int32_t timesteps = static_cast<int32_t>(src->numel_) / vocab;
-        const bool softmax = opts.getProperty(rt, "softmax").asBool();
-
-        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
-        if (!srcLock.owns_lock()) {
-            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor is currently in use");
-        }
-        if (!src->data_) {
-            throw jsi::JSError(rt, "ctcGreedyDecode: src tensor has been disposed");
-        }
         const auto *data = reinterpret_cast<const float *>(src->data_.get());
 
         jsi::Array out(rt, static_cast<size_t>(timesteps) * 2);
@@ -668,23 +648,14 @@ void install_ctcGreedyDecode(jsi::Runtime &rt, jsi::Object &module) {
             const float *row = data + static_cast<std::size_t>(t) * static_cast<std::size_t>(vocab);
             const float *maxIt = std::max_element(row, row + vocab);
             const auto maxIdx = static_cast<int32_t>(maxIt - row);
-            const float maxVal = *maxIt;
-            auto prob = static_cast<double>(maxVal);
-            if (softmax) {
-                double sum = 0.0;
-                for (int32_t v = 0; v < vocab; ++v) {
-                    sum += std::exp(static_cast<double>(row[v]) - static_cast<double>(maxVal));
-                }
-                prob = sum > 0.0 ? 1.0 / sum : 0.0; // exp(maxVal - maxVal) / sum
-            }
             out.setValueAtIndex(rt, oi++, jsi::Value(static_cast<double>(maxIdx)));
-            out.setValueAtIndex(rt, oi++, jsi::Value(prob));
+            out.setValueAtIndex(rt, oi++, jsi::Value(static_cast<double>(*maxIt)));
         }
         return out;
     };
     module.setProperty(rt, name,
                        jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
-                                                             2, fnBody));
+                                                             1, fnBody));
 }
 
 } // namespace rnexecutorch::extensions::cv::ocr_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
index 893123fec3..002fb64601 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
@@ -3,11 +3,16 @@
 #include <jsi/jsi.h>
 
 namespace rnexecutorch::extensions::cv::ocr_ops {
-// Detector heatmap -> flat array of oriented quads (10 doubles/box:
-// x0,y0..x3,y3,score,angle), in detector-input pixels.
-void install_extractTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+// CRAFT half-res region+affinity heatmap [..,Hd,Wd,2] -> flat array of oriented
+// quads (10 doubles/box: x0,y0..x3,y3,score,angle), in detector-input pixels.
+void install_extractCraftTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 
-// Per-timestep argmax + max value over [..,T,V] logits -> flat [idx,prob,...];
-// options.softmax makes `prob` a probability.
+// DBNet full-res probability map [..,H,W] -> flat array of oriented quads (10
+// doubles/box), in detector-input pixels.
+void install_extractDbnetTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
+
+// Per-timestep argmax + max value over [..,T,V] logits -> flat [idx,value,...].
+// `value` is the raw max activation; softmax the tensor beforehand if a caller
+// needs a probability.
 void install_ctcGreedyDecode(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::ocr_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
deleted file mode 100644
index 59f1e45ff6..0000000000
--- a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.cpp
+++ /dev/null
@@ -1,266 +0,0 @@
-#include "text_boxes_ops.h"
-
-#include <algorithm>
-#include <array>
-#include <cmath>
-#include <numeric>
-#include <stdexcept>
-#include <utility>
-
-#include <opencv2/imgproc.hpp>
-
-#include "core/dtype.h"
-#include "core/tensor.h"
-#include "utils.h"
-namespace rnexecutorch::extensions::cv::text_boxes_ops {
-
-namespace jsi = facebook::jsi;
-using TensorHostObject = rnexecutorch::core::tensor::TensorHostObject;
-
-void install_gridSample(jsi::Runtime &rt, jsi::Object &module) {
-    const auto *name = "gridSample";
-    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args, size_t count) -> jsi::Value {
-        if (count != 3) {
-            throw jsi::JSError(rt, "Usage: gridSample(src, grid, dst)");
-        }
-        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt) ||
-            !args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt) ||
-            !args[2].isObject() || !args[2].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "gridSample: src, grid, and dst must be Tensors");
-        }
-
-        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto grid = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto dst = args[2].asObject(rt).getHostObject<TensorHostObject>(rt);
-
-        if (src.get() == dst.get()) {
-            throw jsi::JSError(rt, "gridSample: In-place operations (src == dst) are not supported.");
-        }
-        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
-            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
-            throw jsi::JSError(rt, "gridSample: src and dst must be uint8");
-        }
-        if (grid->dtype_ != rnexecutorch::core::types::DType::float32) {
-            throw jsi::JSError(rt, "gridSample: grid must be float32");
-        }
-        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
-            throw jsi::JSError(rt, "gridSample: src and dst must be [H, W, C]");
-        }
-        if (src->shape_ != dst->shape_) {
-            throw jsi::JSError(rt, "gridSample: src and dst must have the same shape");
-        }
-        // grid is the torch grid_sample field [..,2,gH,gW], channel 0 = x, 1 = y,
-        // normalized to [-1,1] with align_corners=true.
-        const auto &gs = grid->shape_;
-        if (gs.size() < 3 || gs[gs.size() - 3] != 2) {
-            throw jsi::JSError(rt, "gridSample: grid must be [..,2,gH,gW]");
-        }
-
-        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
-        std::shared_lock<std::shared_mutex> gridLock(grid->mutex_, std::try_to_lock);
-        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
-        if (!srcLock.owns_lock() || !gridLock.owns_lock() || !dstLock.owns_lock()) {
-            throw jsi::JSError(rt, "gridSample: a tensor is currently in use");
-        }
-        if (!src->data_ || !grid->data_ || !dst->data_) {
-            throw jsi::JSError(rt, "gridSample: a tensor has been disposed");
-        }
-
-        const int32_t h = src->shape_[0];
-        const int32_t w = src->shape_[1];
-        const int32_t channels = src->shape_[2];
-        const int32_t gridH = gs[gs.size() - 2];
-        const int32_t gridW = gs[gs.size() - 1];
-        const int32_t plane = gridH * gridW;
-        const auto *g = reinterpret_cast<const float *>(grid->data_.get());
-
-        // Bilinearly sample channel `c` of the low-res grid at fractional (gx, gy).
-        auto sampleGrid = [&](int32_t c, float gx, float gy) -> float {
-            const int32_t x0 = std::clamp(static_cast<int32_t>(std::floor(gx)), 0, gridW - 1);
-            const int32_t y0 = std::clamp(static_cast<int32_t>(std::floor(gy)), 0, gridH - 1);
-            const int32_t x1 = std::min(x0 + 1, gridW - 1);
-            const int32_t y1 = std::min(y0 + 1, gridH - 1);
-            const float dx = gx - static_cast<float>(x0);
-            const float dy = gy - static_cast<float>(y0);
-            const int32_t base = c * plane;
-            const float top = g[base + y0 * gridW + x0] +
-                              (g[base + y0 * gridW + x1] - g[base + y0 * gridW + x0]) * dx;
-            const float bot = g[base + y1 * gridW + x0] +
-                              (g[base + y1 * gridW + x1] - g[base + y1 * gridW + x0]) * dx;
-            return top + (bot - top) * dy;
-        };
-
-        ::cv::Mat mapX(h, w, CV_32F);
-        ::cv::Mat mapY(h, w, CV_32F);
-        for (int32_t oy = 0; oy < h; ++oy) {
-            const float gy = h > 1 ? (static_cast<float>(oy) / static_cast<float>(h - 1)) *
-                                         static_cast<float>(gridH - 1)
-                                   : 0.0f;
-            auto *rowX = mapX.ptr<float>(oy);
-            auto *rowY = mapY.ptr<float>(oy);
-            for (int32_t ox = 0; ox < w; ++ox) {
-                const float gx = w > 1 ? (static_cast<float>(ox) / static_cast<float>(w - 1)) *
-                                             static_cast<float>(gridW - 1)
-                                       : 0.0f;
-                const float nx = sampleGrid(0, gx, gy); // [-1,1]
-                const float ny = sampleGrid(1, gx, gy);
-                rowX[ox] = ((nx + 1.0f) / 2.0f) * static_cast<float>(w - 1);
-                rowY[ox] = ((ny + 1.0f) / 2.0f) * static_cast<float>(h - 1);
-            }
-        }
-
-        const int cvType = CV_MAKETYPE(CV_8U, channels);
-        ::cv::Mat srcMat(h, w, cvType, src->data_.get());
-        ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
-        try {
-            ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
-        } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("gridSample: OpenCV error: ") + e.what());
-        }
-        return jsi::Value(rt, args[2]);
-    };
-    module.setProperty(rt, name, jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name), 3, fnBody));
-}
-
-// ------------------------------- warpQuad ----------------------------------
-// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
-// resize-to-height + pad/align). A generic image op; used by the OCR recognizer.
-void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
-    const auto *name = "warpQuad";
-    auto fnBody = [](jsi::Runtime &rt, const jsi::Value &, const jsi::Value *args,
-                     size_t count) -> jsi::Value {
-        if (count != 4) {
-            throw jsi::JSError(rt, "Usage: warpQuad(src, dst, quad, options)");
-        }
-        if (!args[0].isObject() || !args[0].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: src must be a Tensor");
-        }
-        if (!args[1].isObject() || !args[1].asObject(rt).isHostObject<TensorHostObject>(rt)) {
-            throw jsi::JSError(rt, "warpQuad: dst must be a Tensor");
-        }
-        if (!args[2].isObject() || !args[2].asObject(rt).isArray(rt)) {
-            throw jsi::JSError(rt, "warpQuad: quad must be an array of 8 numbers");
-        }
-        if (!args[3].isObject()) {
-            throw jsi::JSError(rt, "warpQuad: options must be an object");
-        }
-        auto src = args[0].asObject(rt).getHostObject<TensorHostObject>(rt);
-        auto dst = args[1].asObject(rt).getHostObject<TensorHostObject>(rt);
-        if (src.get() == dst.get()) {
-            throw jsi::JSError(rt, "warpQuad: In-place operations (src == dst) are not supported.");
-        }
-        auto quadArr = args[2].asObject(rt).asArray(rt);
-        auto opts = args[3].asObject(rt);
-
-        if (quadArr.length(rt) != 8) {
-            throw jsi::JSError(rt, "warpQuad: quad must have exactly 8 numbers (4 points)");
-        }
-        if (src->shape_.size() != 3 || dst->shape_.size() != 3) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be [H,W,C]");
-        }
-        if (src->dtype_ != rnexecutorch::core::types::DType::uint8 ||
-            dst->dtype_ != rnexecutorch::core::types::DType::uint8) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must be uint8");
-        }
-        if (src->shape_[2] != dst->shape_[2]) {
-            throw jsi::JSError(rt, "warpQuad: src and dst must have the same channel count");
-        }
-
-        const int32_t channels = src->shape_[2];
-        const int32_t recH = dst->shape_[0];
-        const int32_t bucketW = dst->shape_[1];
-
-        if (!opts.hasProperty(rt, "contentWidth") ||
-            !opts.getProperty(rt, "contentWidth").isNumber()) {
-            throw jsi::JSError(rt, "warpQuad: options.contentWidth is required");
-        }
-        const int32_t contentWidth =
-            std::clamp(static_cast<int32_t>(opts.getProperty(rt, "contentWidth").asNumber()), 1,
-                       bucketW);
-        const std::string padMode = opts.getProperty(rt, "padMode").asString(rt).utf8(rt);
-        const double padValue = opts.getProperty(rt, "padValue").asNumber();
-        const std::string align = opts.getProperty(rt, "align").asString(rt).utf8(rt);
-        // offsetX >= 0 places content at that x (overriding align); clear=false skips
-        // wiping dst first, so successive warps compose into one canvas (glyph strips).
-        const auto offsetXOpt = static_cast<int32_t>(opts.getProperty(rt, "offsetX").asNumber());
-        const bool clear = opts.getProperty(rt, "clear").asBool();
-
-        std::array<::cv::Point2f, 4> quad;
-        for (std::size_t i = 0; i < 8; ++i) {
-            if (!quadArr.getValueAtIndex(rt, i).isNumber()) {
-                throw jsi::JSError(rt, "warpQuad: quad must contain only numbers");
-            }
-        }
-        for (std::size_t i = 0; i < 4; ++i) {
-            quad[i] = {static_cast<float>(quadArr.getValueAtIndex(rt, i * 2).asNumber()),
-                       static_cast<float>(quadArr.getValueAtIndex(rt, i * 2 + 1).asNumber())};
-        }
-
-        std::shared_lock<std::shared_mutex> srcLock(src->mutex_, std::try_to_lock);
-        if (!srcLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: src tensor is currently in use");
-        }
-        std::unique_lock<std::shared_mutex> dstLock(dst->mutex_, std::try_to_lock);
-        if (!dstLock.owns_lock()) {
-            throw jsi::JSError(rt, "warpQuad: dst tensor is currently in use");
-        }
-        if (!src->data_ || !dst->data_) {
-            throw jsi::JSError(rt, "warpQuad: a tensor has been disposed");
-        }
-
-        const int cvType = CV_MAKETYPE(CV_8U, channels);
-        ::cv::Mat srcMat(src->shape_[0], src->shape_[1], cvType, src->data_.get());
-        ::cv::Mat dstMat(recH, bucketW, cvType, dst->data_.get());
-
-        try {
-            const std::array<::cv::Point2f, 4> dstPts = {
-                ::cv::Point2f{0.0f, 0.0f},
-                {static_cast<float>(contentWidth), 0.0f},
-                {static_cast<float>(contentWidth), static_cast<float>(recH)},
-                {0.0f, static_cast<float>(recH)}};
-            const std::array<::cv::Point2f, 4> srcPts = {quad[0], quad[1], quad[2], quad[3]};
-            ::cv::Mat m = ::cv::getPerspectiveTransform(srcPts.data(), dstPts.data());
-            ::cv::Mat content;
-            ::cv::warpPerspective(srcMat, content, m, ::cv::Size(contentWidth, recH),
-                                  ::cv::INTER_CUBIC, ::cv::BORDER_REPLICATE);
-
-            ::cv::Scalar padColor;
-            if (padMode == "cornerMean") {
-                const int patch = std::max(1, std::min(recH, contentWidth) / 30);
-                ::cv::Scalar acc(0, 0, 0, 0);
-                const std::array<::cv::Rect, 4> rects = {
-                    ::cv::Rect(0, 0, patch, patch),
-                    ::cv::Rect(contentWidth - patch, 0, patch, patch),
-                    ::cv::Rect(0, recH - patch, patch, patch),
-                    ::cv::Rect(contentWidth - patch, recH - patch, patch, patch)};
-                for (const auto &r : rects) {
-                    acc += ::cv::mean(content(r));
-                }
-                padColor = acc / 4.0;
-            } else {
-                padColor = ::cv::Scalar::all(padValue);
-            }
-
-            if (clear) {
-                dstMat.setTo(padColor);
-            }
-            int32_t offsetX = offsetXOpt;
-            if (offsetX < 0) {
-                offsetX = (align == "center") ? (bucketW - contentWidth) / 2 : 0;
-            }
-            if (offsetX < bucketW) {
-                const int32_t copyW = std::min(contentWidth, bucketW - offsetX);
-                content(::cv::Rect(0, 0, copyW, recH))
-                    .copyTo(dstMat(::cv::Rect(offsetX, 0, copyW, recH)));
-            }
-        } catch (const ::cv::Exception &e) {
-            throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
-        }
-        return jsi::Value(rt, args[1]);
-    };
-    module.setProperty(rt, name,
-                       jsi::Function::createFromHostFunction(rt, jsi::PropNameID::forAscii(rt, name),
-                                                             4, fnBody));
-}
-
-} // namespace rnexecutorch::extensions::cv::text_boxes_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h b/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h
deleted file mode 100644
index 99ad684e28..0000000000
--- a/packages/react-native-executorch/cpp/extensions/cv/text_boxes_ops.h
+++ /dev/null
@@ -1,8 +0,0 @@
-#pragma once
-
-#include <jsi/jsi.h>
-
-namespace rnexecutorch::extensions::cv::text_boxes_ops {
-void install_gridSample(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-} // namespace rnexecutorch::extensions::cv::text_boxes_ops

From b521f1fa5e6536802b6076ff144e255915b7e5f4 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 00:15:51 +0200
Subject: [PATCH 25/29] [RNE Rewrite] refactor(ocr): Ocr camelcase, pluggable
 extractors, leak-safe factories

---
 .../react-native-executorch/src/constants.ts  |  11 +-
 .../src/extensions/cv/ops/image.ts            |  57 ++
 .../src/extensions/cv/ops/index.ts            |   1 +
 .../src/extensions/cv/ops/quad.ts             | 229 ++++++++
 .../src/extensions/cv/ops/textBoxes.ts        | 206 -------
 .../tasks/{documentOCR.ts => documentOcr.ts}  | 167 +++---
 .../src/extensions/cv/tasks/ocr.ts            | 324 ++++-------
 .../tasks/ocr/{ocrSymbols.ts => charsets.ts}  |   2 +-
 .../src/extensions/cv/tasks/ocr/detectors.ts  | 152 +++++
 .../cv/tasks/ocr/documentHelpers.ts           | 162 +++---
 .../extensions/cv/tasks/ocr/documentModels.ts | 329 +++++++++++
 .../src/extensions/cv/tasks/ocr/ocrHelpers.ts | 542 ------------------
 .../extensions/cv/tasks/ocr/ocrPipeline.ts    | 539 -----------------
 .../src/extensions/cv/tasks/ocr/ocrUtils.ts   | 327 +++++++++++
 .../src/extensions/cv/tasks/ocr/pipeline.ts   | 368 ++++++++++++
 .../src/extensions/cv/tasks/ocr/supporting.ts | 302 ----------
 .../{useDocumentOCR.ts => useDocumentOcr.ts}  |  38 +-
 .../src/hooks/{useOCR.ts => useOcr.ts}        |  16 +-
 packages/react-native-executorch/src/index.ts |   9 +-
 .../react-native-executorch/src/models.ts     |  92 +--
 20 files changed, 1811 insertions(+), 2062 deletions(-)
 create mode 100644 packages/react-native-executorch/src/extensions/cv/ops/quad.ts
 delete mode 100644 packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
 rename packages/react-native-executorch/src/extensions/cv/tasks/{documentOCR.ts => documentOcr.ts} (54%)
 rename packages/react-native-executorch/src/extensions/cv/tasks/ocr/{ocrSymbols.ts => charsets.ts} (99%)
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
 delete mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
 delete mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
 create mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
 delete mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
 rename packages/react-native-executorch/src/hooks/{useDocumentOCR.ts => useDocumentOcr.ts} (58%)
 rename packages/react-native-executorch/src/hooks/{useOCR.ts => useOcr.ts} (75%)

diff --git a/packages/react-native-executorch/src/constants.ts b/packages/react-native-executorch/src/constants.ts
index 45cf3290e2..eda8ac9769 100644
--- a/packages/react-native-executorch/src/constants.ts
+++ b/packages/react-native-executorch/src/constants.ts
@@ -1301,13 +1301,11 @@ export type BlazeFaceLandmark = (typeof BLAZEFACE_LANDMARKS)[number];
  * @category Types
  */
 export type CocoLandmark = (typeof COCO_LANDMARKS)[number];
-export { alphabets, symbols, PPOCR_SYMBOLS } from './extensions/cv/tasks/ocr/ocrSymbols';
-export type { OCRLanguage } from './extensions/cv/tasks/ocr/ocrSymbols';
+export { alphabets, symbols, PPOCR_SYMBOLS } from './extensions/cv/tasks/ocr/charsets';
+export type { OcrLanguage } from './extensions/cv/tasks/ocr/charsets';
 
 /**
- * PP-DocLayoutV3 region classes, in model output order (index = class id). Some
- * display names repeat at different ids — that is the model's own id2label, kept
- * positional so `DOC_LAYOUT_LABELS[id]` matches the logits dimension.
+ * PP-DocLayoutV3 region classes, in model output order (index = class id).
  * @category Constants
  */
 export const DOC_LAYOUT_LABELS = [
@@ -1346,9 +1344,6 @@ export type DocLayoutLabel = (typeof DOC_LAYOUT_LABELS)[number];
 
 /**
  * SLANet_plus table-structure token vocabulary (50 tokens; index = token id).
- * `0:'sos'`, `49:'eos'`. The table decoder emits the HTML `<tr>/<td>` skeleton
- * (with col/row spans) one token at a time; concatenating the non-special tokens
- * yields the table structure HTML.
  * @category Constants
  */
 export const SLANET_STRUCTURE_VOCAB = [
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/image.ts b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
index c7eec8fd8b..1a6cd1ce7f 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/image.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/image.ts
@@ -224,3 +224,60 @@ export function rotate(src: Tensor, dst: Tensor, degCW: number): Tensor {
   'worklet';
   return rnexecutorchJsi.cv.rotate(src, dst, degCW);
 }
+
+/**
+ * Options for {@link warpQuad}. `contentWidth` is the warped content's width (px)
+ * in the canvas; `align` (`'left'`/`'center'`, default `'left'`) with `padMode`
+ * (`'constant'`/`'cornerMean'`, default `'constant'`) and `padValue` (default `0`)
+ * place and fill it. `offsetX` (default `-1` = use `align`) pins the content at an
+ * exact x, and `clear` (default `true`) wipes the canvas first — pass an explicit
+ * `offsetX` with `clear: false` to compose successive warps side-by-side into one
+ * canvas (e.g. a glyph strip).
+ * @category Types
+ */
+export type WarpQuadOptions = {
+  readonly contentWidth: number;
+  readonly align?: 'left' | 'center';
+  readonly padMode?: 'constant' | 'cornerMean';
+  readonly padValue?: number;
+  readonly offsetX?: number;
+  readonly clear?: boolean;
+};
+
+/**
+ * Perspective-crops an oriented quad region of `src` into the pre-allocated canvas
+ * `dst`, folding crop + resize-to-height + pad into one native pass.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
+ * @param dst The pre-allocated destination canvas in HWC uint8 layout.
+ * @param quad Eight numbers `[x0,y0,..,x3,y3]` (TL,TR,BR,BL) in `src` pixels.
+ * @param opts Content width, alignment, and padding configuration.
+ * @returns The destination tensor `dst`.
+ */
+export function warpQuad(src: Tensor, dst: Tensor, quad: number[], opts: WarpQuadOptions): Tensor {
+  'worklet';
+  return rnexecutorchJsi.cv.warpQuad(src, dst, quad, {
+    contentWidth: opts.contentWidth,
+    align: opts.align ?? 'left',
+    padMode: opts.padMode ?? 'constant',
+    padValue: opts.padValue ?? 0,
+    offsetX: opts.offsetX ?? -1,
+    clear: opts.clear ?? true,
+  });
+}
+
+/**
+ * Warps `src` through a backward sampling field (a `torch.grid_sample`-style remap
+ * — the grid gives, per output pixel, where to read from in `src`) into the
+ * pre-allocated `dst`, natively via `cv::remap`.
+ * @category Typescript API
+ * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
+ * @param grid The sampling field tensor (float32), shape `[..,2,gH,gW]`, channel
+ * 0 = x and 1 = y, normalized to `[-1, 1]` with `align_corners=true`.
+ * @param dst The pre-allocated destination tensor, same shape/dtype as `src`.
+ * @returns The destination tensor `dst`.
+ */
+export function warpByGrid(src: Tensor, grid: Tensor, dst: Tensor): Tensor {
+  'worklet';
+  return rnexecutorchJsi.cv.warpByGrid(src, grid, dst);
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/index.ts b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
index 84a274101d..4128d25527 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/index.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/index.ts
@@ -1,3 +1,4 @@
 export * as image from './image';
 export * as boxes from './boxes';
 export * as points from './points';
+export * as quad from './quad';
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
new file mode 100644
index 0000000000..99ebb01daa
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -0,0 +1,229 @@
+import { scalePoint, type Point } from './points';
+
+/**
+ * An oriented quadrilateral in pixel space: `points` are the four corners ordered
+ * top-left, top-right, bottom-right, bottom-left, `score` is the region confidence
+ * in `[0, 1]`, and `angle` is the rotation in degrees.
+ * @category Types
+ */
+export type Quad = {
+  readonly points: readonly Point[];
+  readonly score: number;
+  readonly angle: number;
+};
+
+/**
+ * The axis-aligned bounds of a set of points.
+ * @category Types
+ */
+export type Bounds = { xmin: number; ymin: number; xmax: number; ymax: number };
+
+const distance = (a: Point, b: Point): number => {
+  'worklet';
+  return Math.hypot(b.x - a.x, b.y - a.y);
+};
+
+const lerp = (a: Point, b: Point, t: number): Point => {
+  'worklet';
+  return { x: a.x + (b.x - a.x) * t, y: a.y + (b.y - a.y) * t };
+};
+
+/**
+ * Computes the axis-aligned bounds enclosing a set of points. Returns a zero box
+ * for empty input.
+ * @category Typescript API
+ * @param points The points to enclose.
+ * @returns The enclosing `{ xmin, ymin, xmax, ymax }` bounds.
+ */
+export function boundsOfPoints(points: readonly Point[]): Bounds {
+  'worklet';
+  if (points.length === 0) {
+    return { xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
+  }
+  let xmin = Infinity;
+  let ymin = Infinity;
+  let xmax = -Infinity;
+  let ymax = -Infinity;
+  for (const p of points) {
+    if (p.x < xmin) xmin = p.x;
+    if (p.y < ymin) ymin = p.y;
+    if (p.x > xmax) xmax = p.x;
+    if (p.y > ymax) ymax = p.y;
+  }
+  return { xmin, ymin, xmax, ymax };
+}
+
+/**
+ * Orders four corner points as top-left, top-right, bottom-right, bottom-left
+ * using their coordinate-sum and coordinate-difference extremes. Inputs that do
+ * not have exactly four points are returned unchanged.
+ * @category Typescript API
+ * @param points The four unordered corners.
+ * @returns The corners ordered TL, TR, BR, BL.
+ */
+export function orderQuad(points: readonly Point[]): Point[] {
+  'worklet';
+  if (points.length !== 4) {
+    return [...points];
+  }
+  let topLeft = 0;
+  let topRight = 0;
+  let bottomRight = 0;
+  let bottomLeft = 0;
+  let minSum = points[0]!.x + points[0]!.y;
+  let maxSum = minSum;
+  let minDiff = points[0]!.y - points[0]!.x;
+  let maxDiff = minDiff;
+  for (let i = 1; i < 4; i++) {
+    const sum = points[i]!.x + points[i]!.y;
+    const diff = points[i]!.y - points[i]!.x;
+    if (sum < minSum) {
+      minSum = sum;
+      topLeft = i;
+    }
+    if (sum > maxSum) {
+      maxSum = sum;
+      bottomRight = i;
+    }
+    if (diff < minDiff) {
+      minDiff = diff;
+      topRight = i;
+    }
+    if (diff > maxDiff) {
+      maxDiff = diff;
+      bottomLeft = i;
+    }
+  }
+  return [points[topLeft]!, points[topRight]!, points[bottomRight]!, points[bottomLeft]!];
+}
+
+/**
+ * Computes the width and height (in pixels) of an ordered TL,TR,BR,BL quad, taking
+ * the longer of each pair of opposite sides.
+ * @category Typescript API
+ * @param ordered The quad corners ordered TL, TR, BR, BL.
+ * @returns The quad's width and height in pixels.
+ */
+export function quadSize(ordered: readonly Point[]): { width: number; height: number } {
+  'worklet';
+  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
+  const width = Math.max(distance(tl, tr), distance(bl, br));
+  const height = Math.max(distance(tl, bl), distance(tr, br));
+  return { width, height };
+}
+
+/**
+ * Maps a quad expressed in a resized (letterboxed) frame back to the original
+ * image frame, clamping the result to the image bounds.
+ * @category Typescript API
+ * @param quad The quad in the resized frame.
+ * @param fromWidth The width of the resized frame the quad is expressed in.
+ * @param fromHeight The height of the resized frame the quad is expressed in.
+ * @param toWidth The original image width.
+ * @param toHeight The original image height.
+ * @returns The four corners in original image pixels.
+ */
+export function mapQuadToImage(
+  quad: Quad,
+  fromWidth: number,
+  fromHeight: number,
+  toWidth: number,
+  toHeight: number
+): Point[] {
+  'worklet';
+  return quad.points.map((p) => {
+    const m = scalePoint(p, {
+      from: { width: fromWidth, height: fromHeight },
+      to: { width: toWidth, height: toHeight },
+      resizeMode: 'letterbox',
+    });
+    return { x: Math.max(0, Math.min(m.x, toWidth)), y: Math.max(0, Math.min(m.y, toHeight)) };
+  });
+}
+
+/**
+ * Splits an ordered TL,TR,BR,BL quad into `parts` equal vertical bands (each an
+ * ordered quad), top to bottom. `parts <= 1` returns the quad unchanged.
+ * @category Typescript API
+ * @param ordered The quad corners ordered TL, TR, BR, BL.
+ * @param parts The number of equal vertical bands to split into.
+ * @returns The bands as ordered TL,TR,BR,BL quads, top to bottom.
+ */
+export function splitTallQuad(ordered: readonly Point[], parts: number): Point[][] {
+  'worklet';
+  if (parts <= 1) {
+    return [ordered as Point[]];
+  }
+  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
+  const out: Point[][] = [];
+  for (let i = 0; i < parts; i++) {
+    const t0 = i / parts;
+    const t1 = (i + 1) / parts;
+    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
+  }
+  return out;
+}
+
+/**
+ * Computes the axis-aligned bounding quad (ordered TL,TR,BR,BL) enclosing a set of
+ * quads. Returns a zero quad for empty input.
+ * @category Typescript API
+ * @param quads The quads to enclose.
+ * @returns The four enclosing corners, ordered TL, TR, BR, BL.
+ */
+export function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
+  'worklet';
+  const all: Point[] = [];
+  for (const q of quads) {
+    for (const p of q) {
+      all.push(p);
+    }
+  }
+  const { xmin, ymin, xmax, ymax } = boundsOfPoints(all);
+  return [
+    { x: xmin, y: ymin },
+    { x: xmax, y: ymin },
+    { x: xmax, y: ymax },
+    { x: xmin, y: ymax },
+  ];
+}
+
+/**
+ * Flattens an ordered TL,TR,BR,BL quad into the 8-number `[x0,y0,..,x3,y3]` array.
+ * @category Typescript API
+ * @param corners The four quad corners (TL, TR, BR, BL).
+ * @returns The eight coordinates `[x0,y0,x1,y1,x2,y2,x3,y3]`.
+ */
+export function flattenQuad(corners: readonly Point[]): number[] {
+  'worklet';
+  // prettier-ignore
+  return [
+    corners[0]!.x, corners[0]!.y, corners[1]!.x, corners[1]!.y,
+    corners[2]!.x, corners[2]!.y, corners[3]!.x, corners[3]!.y,
+  ];
+}
+
+/**
+ * Builds oriented quads from a detector's flat output array — 10 numbers per box:
+ * `x0,y0,..,x3,y3,score,angle`.
+ * @category Typescript API
+ * @param flat The flat number array from a native detector decode.
+ * @returns The parsed quads.
+ */
+export function quadsFromFlat(flat: number[]): Quad[] {
+  'worklet';
+  const quads: Quad[] = [];
+  for (let i = 0; i < flat.length; i += 10) {
+    quads.push({
+      points: [
+        { x: flat[i]!, y: flat[i + 1]! },
+        { x: flat[i + 2]!, y: flat[i + 3]! },
+        { x: flat[i + 4]!, y: flat[i + 5]! },
+        { x: flat[i + 6]!, y: flat[i + 7]! },
+      ],
+      score: flat[i + 8]!,
+      angle: flat[i + 9]!,
+    });
+  }
+  return quads;
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
deleted file mode 100644
index 429047f930..0000000000
--- a/packages/react-native-executorch/src/extensions/cv/ops/textBoxes.ts
+++ /dev/null
@@ -1,206 +0,0 @@
-import { rnexecutorchJsi } from '../../../native/bridge';
-import type { Tensor } from '../../../core/tensor';
-import type { Point } from './points';
-
-/**
- * Detector heatmap postprocessing mode. `'craft'` consumes a half-resolution
- * interleaved `[..,Hd,Wd,2]` (text, affinity) map; `'dbnet'` consumes a
- * full-resolution `[..,H,W]` probability map.
- * @category Types
- */
-export type TextBoxMode = 'craft' | 'dbnet';
-
-/**
- * An oriented quadrilateral text region returned by {@link extractTextBoxes},
- * in detector-input pixel space.
- * @category Types
- */
-export type Quad = {
-  /** Four corner points, ordered TL, TR, BR, BL. */
-  readonly points: readonly Point[];
-  /** Detector confidence for this region (1 for CRAFT grouped lines). */
-  readonly score: number;
-  /** Line angle in degrees. */
-  readonly angle: number;
-};
-
-/**
- * CRAFT-mode options for {@link extractTextBoxes} (region+affinity heatmap).
- * @category Types
- */
-export type CraftExtractOptions = {
-  readonly mode: 'craft';
-  readonly textThreshold: number;
-  readonly linkThreshold: number;
-  readonly lowTextThreshold: number;
-  /**
-   * Detector input height — used to restore the half-res boxes. Required (the
-   * native op throws without it); the OCR pipeline sets it per run.
-   */
-  readonly targetHeight: number;
-  /**
-   * Return one upright box per glyph instead of grouped reading-ordered lines:
-   * the affinity map is subtracted (not added) to separate adjacent characters,
-   * and no line grouping or de-skew rotation is applied. Used by the second,
-   * per-column pass that reads upright stacked text. Defaults to `false`.
-   */
-  readonly charLevel?: boolean;
-};
-
-/**
- * DBNet-mode options for {@link extractTextBoxes} (probability map).
- * @category Types
- */
-export type DbnetExtractOptions = {
-  readonly mode: 'dbnet';
-  readonly binThreshold: number;
-  readonly boxThreshold: number;
-  readonly unclipRatio: number;
-  readonly minBoxSide: number;
-  readonly maxCandidates: number;
-  /**
-   * Whether the probability map needs a sigmoid applied first. Set from the
-   * model's export contract (`true` = raw logits, `false` = already probabilities).
-   */
-  readonly applySigmoid: boolean;
-};
-
-/**
- * Options for {@link extractTextBoxes} — a discriminated union on `mode`, so the
- * type system enforces exactly the field set the chosen detector needs.
- * @category Types
- */
-export type ExtractTextBoxesOptions = CraftExtractOptions | DbnetExtractOptions;
-
-/**
- * Converts a detector heatmap tensor into oriented text-line quads.
- *
- * Variable-sized output: the native op returns a flat array of 10 doubles per
- * box which this wrapper reshapes into {@link Quad}s.
- * @category Typescript API
- * @param src The detector output tensor (float32). CRAFT: `[..,Hd,Wd,2]`;
- * DBNet: `[..,H,W]`.
- * @param opts Mode and per-mode thresholds.
- * @returns The detected quads in detector-input pixel space.
- */
-export function extractTextBoxes(src: Tensor, opts: ExtractTextBoxesOptions): Quad[] {
-  'worklet';
-  const flat = rnexecutorchJsi.cv.extractTextBoxes(src, opts) as number[];
-  const quads: Quad[] = [];
-  for (let i = 0; i < flat.length; i += 10) {
-    quads.push({
-      points: [
-        { x: flat[i]!, y: flat[i + 1]! },
-        { x: flat[i + 2]!, y: flat[i + 3]! },
-        { x: flat[i + 4]!, y: flat[i + 5]! },
-        { x: flat[i + 6]!, y: flat[i + 7]! },
-      ],
-      score: flat[i + 8]!,
-      angle: flat[i + 9]!,
-    });
-  }
-  return quads;
-}
-
-/**
- * Options for {@link warpQuad}.
- * @category Types
- */
-export type WarpQuadOptions = {
-  /** Width (px) of the warped content inside the destination canvas. */
-  readonly contentWidth: number;
-  /** Horizontal placement of the content in the canvas. Default `'left'`. */
-  readonly align?: 'left' | 'center';
-  /** How the remaining canvas is filled. Default `'constant'`. */
-  readonly padMode?: 'constant' | 'cornerMean';
-  /** Fill value for `padMode: 'constant'`. Default `0`. */
-  readonly padValue?: number;
-  /**
-   * Exact x (px) to place the content at, overriding `align`. Default `-1`
-   * (use `align`). Combined with `clear: false`, lets successive warps compose
-   * side-by-side into one canvas (e.g. a glyph strip).
-   */
-  readonly offsetX?: number;
-  /**
-   * Wipe the canvas to the pad color before writing. Default `true`. Set `false`
-   * to preserve prior content (compose multiple warps into one `dst`).
-   */
-  readonly clear?: boolean;
-};
-
-/**
- * Perspective-crops an oriented quad region of `src` into the pre-allocated
- * recognizer canvas `dst`, folding crop + resize-to-height + bucket-pad.
- * @category Typescript API
- * @param src The source image tensor in HWC uint8 layout. Shape [H,W,C].
- * @param dst The pre-allocated recognizer canvas in HWC uint8 layout, sized
- * `[recH, bucketW, C]`.
- * @param quad Eight numbers `[x0,y0,..,x3,y3]` (TL,TR,BR,BL) in `src` pixels.
- * @param opts Content width, alignment, and padding configuration.
- * @returns The destination tensor `dst`.
- */
-export function warpQuad(src: Tensor, dst: Tensor, quad: number[], opts: WarpQuadOptions): Tensor {
-  'worklet';
-  return rnexecutorchJsi.cv.warpQuad(src, dst, quad, {
-    contentWidth: opts.contentWidth,
-    align: opts.align ?? 'left',
-    padMode: opts.padMode ?? 'constant',
-    padValue: opts.padValue ?? 0,
-    offsetX: opts.offsetX ?? -1,
-    clear: opts.clear ?? true,
-  });
-}
-
-/**
- * Options for {@link ctcGreedyDecode}.
- * @category Types
- */
-export type CtcGreedyDecodeOptions = {
-  /**
-   * Apply a per-timestep softmax so `values` are probabilities (CRNN logits).
-   * Leave `false` for heads that are already softmaxed (SVTR).
-   */
-  readonly softmax?: boolean;
-};
-
-/**
- * Per-timestep argmax + max value over recognizer logits, computed natively on
- * the tensor buffer (avoids copying the whole `[T, V]` tensor into JS).
- * @category Typescript API
- * @param src The recognizer output tensor (float32), shape `[..,T,V]`.
- * @param opts Whether to softmax each timestep before taking the max value.
- * @returns The per-timestep argmax `indices` and (optionally softmaxed) max
- * `values`.
- */
-export function ctcGreedyDecode(
-  src: Tensor,
-  opts?: CtcGreedyDecodeOptions
-): { indices: number[]; values: number[] } {
-  'worklet';
-  const flat = rnexecutorchJsi.cv.ctcGreedyDecode(src, {
-    softmax: opts?.softmax ?? false,
-  }) as number[];
-  const indices: number[] = [];
-  const values: number[] = [];
-  for (let i = 0; i < flat.length; i += 2) {
-    indices.push(flat[i]!);
-    values.push(flat[i + 1]!);
-  }
-  return { indices, values };
-}
-
-/**
- * Resamples `src` through a backward sampling field (the `torch.grid_sample`
- * step of a geometric dewarp) into the pre-allocated `dst`, natively via
- * `cv::remap`.
- * @category Typescript API
- * @param src The source image tensor in HWC uint8 layout, shape `[H, W, C]`.
- * @param grid The sampling field tensor (float32), shape `[..,2,gH,gW]`, channel
- * 0 = x and 1 = y, normalized to `[-1, 1]` with `align_corners=true`.
- * @param dst The pre-allocated destination tensor, same shape/dtype as `src`.
- * @returns The destination tensor `dst`.
- */
-export function gridSample(src: Tensor, grid: Tensor, dst: Tensor): Tensor {
-  'worklet';
-  return rnexecutorchJsi.cv.gridSample(src, grid, dst);
-}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
similarity index 54%
rename from packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
index 12d0bb98df..d153162514 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOCR.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
@@ -6,14 +6,14 @@ import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
 import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
 import { rotate, FORMAT_CHANNELS } from '../ops/image';
-import { createOCR, type OCRModel, type OCRDetection } from './ocr';
+import { createOcr, type OcrModel, type OcrDetection } from './ocr';
 import {
   createObjectDetector,
   type ObjectDetectorModel,
   type ObjectDetection,
 } from './objectDetection';
-import { createSupporting, type SupportingModel } from './ocr/supporting';
-import { readingOrderIndices } from './ocr/ocrHelpers';
+import { createDocumentModels, type DocumentModelsConfig } from './ocr/documentModels';
+import { orderByReadingOrder } from './ocr/ocrUtils';
 import { cropImageBuffer, fillTableCells } from './ocr/documentHelpers';
 
 /**
@@ -31,7 +31,7 @@ export type DocumentBlock<L> = {
   /** The block's text, lines joined top-to-bottom by newlines. */
   readonly text: string;
   /** The OCR lines inside this block, top-to-bottom. */
-  readonly lines: readonly OCRDetection[];
+  readonly lines: readonly OcrDetection[];
   /** Whether this block is a table region. */
   readonly isTable: boolean;
   /** For table blocks: the recognized HTML structure with OCR text filled in. */
@@ -45,7 +45,7 @@ export type DocumentBlock<L> = {
 export type DocumentResult<L> = {
   readonly blocks: DocumentBlock<L>[];
   readonly regions: ObjectDetection<'xyxy', L>[];
-  readonly detections: OCRDetection[];
+  readonly detections: OcrDetection[];
   /**
    * The frame all `bbox`/`quad` coordinates are relative to. Equals the input
    * image unless orientation correction or dewarp was applied, in which case it
@@ -56,76 +56,61 @@ export type DocumentResult<L> = {
 
 /**
  * Configuration for the document OCR orchestrator. Provides an OCR model, an
- * optional layout model (regions/blocks), and an optional supporting model
+ * optional layout model (regions/blocks), and optional document models
  * (orientation/dewarp pre-processing + table-structure recognition). The
  * `orientation`/`dewarp` flags are *defaults* for the per-run options of the
  * same name — supply them here to bias every run, or leave them off and pass
- * them to `runDocumentOCR` per call (the supporting model is loaded either way).
+ * them to `runDocumentOcr` per call (the document models are loaded either way).
  * @category Types
  */
-export type DocumentOCRModel<L> = {
-  readonly ocr: OCRModel;
+export type DocumentOcrModel<L> = {
+  readonly ocr: OcrModel;
   readonly layout?: ObjectDetectorModel<'xyxy', L>;
-  readonly supporting?: SupportingModel;
-  /** Default for the per-run `orientation` option (needs `supporting`). */
+  readonly documentModels?: DocumentModelsConfig;
+  /** Default for the per-run `orientation` option (needs `documentModels`). */
   readonly orientation?: boolean;
-  /** Default for the per-run `dewarp` option (needs `supporting`). */
+  /** Default for the per-run `dewarp` option (needs `documentModels`). */
   readonly dewarp?: boolean;
+  /**
+   * Minimum orientation-classifier confidence (softmax of the argmax class) to act
+   * on a non-zero rotation — below it the page is treated as already upright, so
+   * out-of-distribution inputs (photos/non-documents) don't spuriously flip. Genuine
+   * documents score >0.95; defaults to 0.85.
+   */
+  readonly orientationMinConfidence?: number;
 };
 
 /**
- * Per-run document options (passed to `runDocumentOCR`, not baked into the
+ * Per-run document options (passed to `runDocumentOcr`, not baked into the
  * model — toggling them needs no reload). Each pre-processing pass still
- * requires the supporting model to have been loaded (`config.supporting`).
+ * requires the document models to have been loaded (`config.documentModels`).
  * @category Types
  */
-export type RunDocumentOCROptions = {
+export type RunDocumentOcrOptions = {
   /**
-   * Detect + correct page orientation before OCR. No-op without a loaded
-   * supporting model. Defaults to the model's `config.orientation`.
+   * Detect + correct page orientation before OCR. No-op without loaded document
+   * models. Defaults to the model's `config.orientation`.
    */
   readonly orientation?: boolean;
   /**
-   * Geometrically dewarp the page before OCR. No-op without a loaded supporting
-   * model. Defaults to the model's `config.dewarp`.
+   * Geometrically dewarp the page before OCR. No-op without loaded document
+   * models. Defaults to the model's `config.dewarp`.
    */
   readonly dewarp?: boolean;
 };
 
-// Minimum orientation-classifier confidence (softmax of the argmax class) to act
-// on a non-zero rotation. Mirrors PaddleOCR's pipeline gate: out-of-distribution
-// inputs (photos, non-documents) produce low-confidence argmaxes that spuriously
-// flip the page, so below this we treat the page as already upright (0°). Set high
-// (0.85) — genuine documents score >0.95, leaving margin to reject OOD frames that
-// can still land ~0.74.
-const ORIENTATION_MIN_CONFIDENCE = 0.85;
-
 // Layout classes that carry no text — skip OCR on them.
 const VISUAL_LABELS = ['image', 'chart', 'seal'];
-const isTextRegion = (label: unknown): boolean => {
-  'worklet';
-  return !VISUAL_LABELS.includes(String(label));
-};
-
-// Shifts a crop-space detection back into page coordinates.
-function offsetDetection(d: OCRDetection, dx: number, dy: number): OCRDetection {
-  'worklet';
-  return { ...d, quad: d.quad.map((p) => ({ x: p.x + dx, y: p.y + dy })) };
-}
 
 function makeBlock<L>(
   regionType: L | 'ungrouped',
   bbox: BoundingBox<'xyxy'>,
   score: number,
-  lines: OCRDetection[],
+  lines: OcrDetection[],
   isTable: boolean
 ): DocumentBlock<L> {
   'worklet';
-  // Order the block's lines in reading order (top-to-bottom, and left-to-right
-  // within a line) so multi-column regions, titles split into words, and
-  // label/value rows concatenate correctly — not in the detector's arbitrary order.
-  const order = readingOrderIndices(lines.map((l) => l.quad));
-  const sorted = order.map((i) => lines[i]!);
+  const sorted = orderByReadingOrder(lines);
   return {
     regionType,
     bbox,
@@ -138,56 +123,60 @@ function makeBlock<L>(
 
 /**
  * Creates the document OCR orchestrator. Pipeline: correct orientation → dewarp
- * (supporting) → layout → per-region OCR (each text region is cropped and OCR'd on
+ * (document models) → layout → per-region OCR (each text region is cropped and OCR'd on
  * its own, upscaled into the detector — far better recall on dense pages than one
  * whole-page pass; lines are offset back to page coords) → tables recognize their
  * structure and fill cells with that region's OCR. Visual regions are skipped.
- * Without layout it OCRs the whole page into one block. Layout/supporting optional.
+ * Without layout it OCRs the whole page into one block. Layout/document models optional.
  * @category Typescript API
- * @param config OCR model + optional layout + optional supporting + flags.
+ * @param config OCR model + optional layout + optional document models + flags.
  * @param runtime Optional worklet runtime thread.
  * @returns A promise resolving to run + disposal controls.
  */
-export async function createDocumentOCR<L>(
-  config: DocumentOCRModel<L>,
+export async function createDocumentOcr<L>(
+  config: DocumentOcrModel<L>,
   runtime?: WorkletRuntime
 ): Promise<{
   dispose: () => void;
-  runDocumentOCR: (
+  runDocumentOcr: (
     input: ImageBuffer,
-    options?: RunDocumentOCROptions
+    options?: RunDocumentOcrOptions
   ) => Promise<DocumentResult<L>>;
-  runDocumentOCRWorklet: (input: ImageBuffer, options?: RunDocumentOCROptions) => DocumentResult<L>;
+  runDocumentOcrWorklet: (input: ImageBuffer, options?: RunDocumentOcrOptions) => DocumentResult<L>;
 }> {
-  const ocr = await createOCR(config.ocr, runtime);
-  const layout = config.layout
-    ? await createObjectDetector<'xyxy', L>(config.layout, runtime)
-    : null;
-  const supporting = config.supporting ? await createSupporting(config.supporting, runtime) : null;
-  // Per-run orientation/dewarp default to the model's config flags; both are
-  // no-ops without a loaded supporting model.
+  const ocr = await createOcr(config.ocr, runtime);
+  let layout: Awaited<ReturnType<typeof createObjectDetector<'xyxy', L>>> | null = null;
+  let documentModels: Awaited<ReturnType<typeof createDocumentModels>> | null = null;
+  try {
+    layout = config.layout ? await createObjectDetector<'xyxy', L>(config.layout, runtime) : null;
+    documentModels = config.documentModels
+      ? await createDocumentModels(config.documentModels, runtime)
+      : null;
+  } catch (e) {
+    // A later model failing to build must not leak the ones already built.
+    layout?.dispose();
+    ocr.dispose();
+    throw e;
+  }
   const defaultOrientation = !!config.orientation;
   const defaultDewarp = !!config.dewarp;
+  const minConfidence = config.orientationMinConfidence ?? 0.85;
 
   const dispose = () => {
     ocr.dispose();
     layout?.dispose();
-    supporting?.dispose();
+    documentModels?.dispose();
   };
 
-  const runDocumentOCRWorklet = (
+  const runDocumentOcrWorklet = (
     input: ImageBuffer,
-    options?: RunDocumentOCROptions
+    options?: RunDocumentOcrOptions
   ): DocumentResult<L> => {
     'worklet';
-    const useOrientation = !!supporting && (options?.orientation ?? defaultOrientation);
-    const useDewarp = !!supporting && (options?.dewarp ?? defaultDewarp);
+    const useOrientation = !!documentModels && (options?.orientation ?? defaultOrientation);
+    const useDewarp = !!documentModels && (options?.dewarp ?? defaultDewarp);
     let img = input;
-    // Orientation + dewarp thread one page tensor (built once): rotate and the
-    // dewarp remap run tensor -> tensor with no intermediate ImageBuffer, and the
-    // corrected page is materialized back to an ImageBuffer once for the OCR/crop
-    // stages below (which are all ImageBuffer-based).
-    if ((useOrientation || useDewarp) && supporting) {
+    if ((useOrientation || useDewarp) && documentModels) {
       const ch = FORMAT_CHANNELS[input.format];
       let page = tensor('uint8', [input.height, input.width, ch]);
       page.setData(input.data);
@@ -195,12 +184,9 @@ export async function createDocumentOCR<L>(
       let ph = input.height;
       try {
         if (useOrientation) {
-          // Only correct when the classifier is confident AND the predicted angle
-          // is non-zero — a low-confidence argmax (typical of OOD photos / non-
-          // documents) otherwise spuriously flips the page.
-          const orientation = supporting.detectOrientationWorklet(page, input.format);
+          const orientation = documentModels.detectOrientationWorklet(page, input.format);
           const deg = ((360 - orientation.rotationCW) % 360) as 0 | 90 | 180 | 270;
-          if (deg !== 0 && orientation.confidence >= ORIENTATION_MIN_CONFIDENCE) {
+          if (deg !== 0 && orientation.confidence >= minConfidence) {
             const swap = deg === 90 || deg === 270;
             const rotated = tensor('uint8', [swap ? pw : ph, swap ? ph : pw, ch]);
             try {
@@ -218,7 +204,7 @@ export async function createDocumentOCR<L>(
         }
         if (useDewarp) {
           // dewarp returns the input tensor unchanged when it declines the warp.
-          const dewarped = supporting.dewarpWorklet(page, input.format);
+          const dewarped = documentModels.dewarpWorklet(page, input.format);
           if (dewarped !== page) {
             page.dispose();
             page = dewarped;
@@ -232,14 +218,9 @@ export async function createDocumentOCR<L>(
       }
     }
 
-    // OCR runs once per region here (potentially many), so don't let each call
-    // free+reload its bucket arenas (release: false). Instead free the model's
-    // bucket methods ONCE in the finally below, after the whole page — keeping
-    // the page's working set cached while still bounding memory across pages.
     try {
-      // Mode A — no layout: OCR the whole page into one block.
       if (!layout) {
-        const detections = ocr.runOCRWorklet(img, { release: false }).detections;
+        const detections = ocr.runOcrWorklet(img, { release: false }).detections;
         const blocks = detections.length
           ? [
               makeBlock<L>(
@@ -254,21 +235,19 @@ export async function createDocumentOCR<L>(
         return { blocks, regions: [], detections, image: img };
       }
 
-      // Mode B — layout: OCR each text region's crop on its own (upscaled into the
-      // detector → far better recall than one whole-page pass), offsetting lines
-      // back to page coords. Tables also recognize structure + fill cells.
       const regions = layout.detectObjectsWorklet(img);
       const blocks: DocumentBlock<L>[] = [];
-      const detections: OCRDetection[] = [];
+      const detections: OcrDetection[] = [];
       for (const region of regions) {
-        if (!isTextRegion(region.label)) {
+        if (VISUAL_LABELS.includes(String(region.label))) {
           continue;
         }
         const { xmin, ymin } = region.box;
         const crop = cropImageBuffer(img, region.box);
-        const lines = ocr
-          .runOCRWorklet(crop, { release: false })
-          .detections.map((d) => offsetDetection(d, xmin, ymin));
+        const lines = ocr.runOcrWorklet(crop, { release: false }).detections.map((d) => ({
+          ...d,
+          quad: d.quad.map((p) => ({ x: p.x + xmin, y: p.y + ymin })),
+        }));
         if (lines.length === 0 && region.label !== 'table') {
           continue;
         }
@@ -280,8 +259,8 @@ export async function createDocumentOCR<L>(
           lines,
           region.label === 'table'
         );
-        if (region.label === 'table' && supporting) {
-          const structure = supporting.recognizeTableWorklet(crop);
+        if (region.label === 'table' && documentModels) {
+          const structure = documentModels.recognizeTableWorklet(crop);
           block = { ...block, tableHtml: fillTableCells(structure.html, block.lines) };
         }
         blocks.push(block);
@@ -289,13 +268,11 @@ export async function createDocumentOCR<L>(
       blocks.sort((a, b) => a.bbox.ymin - b.bbox.ymin || a.bbox.xmin - b.bbox.xmin);
       return { blocks, regions, detections, image: img };
     } finally {
-      // Free the OCR model's bucket arenas once, after the whole page (the
-      // per-region runs used release: false). Bounds memory across pages while
-      // keeping each page's working set cached during the run.
+      // Per-region runs pass release: false; the bucket arenas are freed once per page.
       ocr.releaseMethodsWorklet();
     }
   };
 
-  const runDocumentOCR = wrapAsync(runDocumentOCRWorklet, runtime);
-  return { runDocumentOCR, runDocumentOCRWorklet, dispose };
+  const runDocumentOcr = wrapAsync(runDocumentOcrWorklet, runtime);
+  return { runDocumentOcr, runDocumentOcrWorklet, dispose };
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 26498c6833..f4d9d298ae 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -6,28 +6,22 @@ import { wrapAsync } from '../../../core/runtime';
 
 import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
-import { FORMAT_CHANNELS, cvtColor } from '../ops/image';
-import type { Quad } from '../ops/textBoxes';
+import { FORMAT_CHANNELS, FORMAT_CONVERSION, cvtColor } from '../ops/image';
+import { orderQuad, quadSize, boundingQuadOf } from '../ops/quad';
+import type { TextBoxExtractor } from './ocr/detectors';
 import {
   buildCharset,
-  orderQuad,
-  quadSize,
-  nowMs,
-  readingOrderIndices,
-  boundingQuadOf,
+  orderByReadingOrder,
   groupVerticalColumns,
   type Buckets,
-} from './ocr/ocrHelpers';
+} from './ocr/ocrUtils';
 import {
-  toRgbCode,
   detectQuads,
   recognizeQuad,
   recognizeGlyphStrip,
-  readBoxVertical,
-  validateDetectorSchema,
-  buildExtractOpts,
-  deriveDetectorOutputs,
-  deriveRecognizerContract,
+  readStackedColumn,
+  resolveDetectorContract,
+  resolveRecognizerContract,
   disposeDetSets,
   disposeRecSets,
   type DetSet,
@@ -35,28 +29,22 @@ import {
   type DetectContext,
   type RecContext,
   type VerticalContext,
-  type DetectorExtractConfig,
-} from './ocr/ocrPipeline';
+} from './ocr/pipeline';
 
-export type { Buckets } from './ocr/ocrHelpers';
+export type { Buckets } from './ocr/ocrUtils';
+export type { Quad } from '../ops/quad';
+export type { TextBoxExtractor } from './ocr/detectors';
 
 /**
- * Configuration for the unified OCR pipeline. A model declares its detector
- * architecture, its input-size buckets, and its charset; the detector/recognizer
- * share one baked contract whose defaults match CRAFT (EasyOCR) and DBNet
- * (PaddleOCR). Models that diverge can override the recognizer normalization,
- * padding, and decode, or supply a `'custom'` detector with its own box
- * extraction — see the per-field options below.
+ * Configuration for the OCR pipeline: a model declares its input-size buckets, its
+ * charset, and its detector box-extraction strategy. The pipeline is
+ * architecture-agnostic — it validates the detect/recognize contract at load and
+ * takes everything model-specific here. The built-in {@link craftExtractBoxes} /
+ * {@link dbnetExtractBoxes} cover EasyOCR / PaddleOCR; other models supply their own
+ * {@link TextBoxExtractor} and override the recognizer normalization/padding/decode.
  * @category Types
  */
-export type OCROptions = {
-  /**
-   * Detector architecture — selects the box decoder (CRAFT heatmap grouping vs
-   * DBNet prob-map) and the default drop score. Use `'custom'` for any other
-   * architecture and supply {@link OCROptions.extractBoxes} to turn the raw
-   * detector output into quads in TypeScript.
-   */
-  readonly detectorKind: 'craft' | 'dbnet' | 'custom';
+export type OcrOptions = {
   /**
    * The model's static input-size buckets. The pipeline snaps each image to the
    * closest `detect`/`recognize` bucket and calls the matching per-size method
@@ -68,36 +56,32 @@ export type OCROptions = {
    * verbatim, for multi-codepoint entries like ligatures).
    */
   readonly charset: string | readonly string[];
-  /** Drop detections below this confidence. Defaults per detector architecture. */
-  readonly dropScore?: number;
   /**
-   * Custom detector post-processing, required when `detectorKind === 'custom'`.
-   * Receives the raw `detect_<S>` output tensors (the model's declared outputs,
-   * in order — shapes read from the PTE, allocated for you) and the snapped
-   * square side `s`, and returns oriented quads in DETECTOR space (the `s × s`
-   * letterboxed input); the pipeline maps them to image pixels and applies
-   * dropScore. Ignored for the built-in kinds. MUST be a worklet — it runs on
-   * the pipeline's worklet thread.
+   * Detector box-extraction strategy: maps the raw `detect_<S>` outputs to oriented
+   * quads. Use the built-in {@link craftExtractBoxes} / {@link dbnetExtractBoxes}, or
+   * supply your own {@link TextBoxExtractor} to plug in a new detector.
    */
-  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
+  readonly extractBoxes: TextBoxExtractor;
+  /** Drop detections scoring below this. Defaults to 0. */
+  readonly dropScore?: number;
   /**
    * Recognizer input normalization, applied after the warp as `x·alpha + beta`
    * (scalar, or per-RGB-channel `[r,g,b]`). Defaults to `(x/255 − 0.5)/0.5` →
-   * `[−1,1]` (`alpha = 1/127.5`, `beta = −1`), the SVTR/CRNN convention. Override
-   * for a recognizer trained with different normalization (e.g. ImageNet).
+   * `[−1,1]` (`alpha = 1/127.5`, `beta = −1`). Override for a recognizer trained
+   * with different normalization (e.g. ImageNet). The detector input norm is
+   * fixed by contract: RGB ÷ 255, with mean/std baked into the PTE.
    */
   readonly recognizerNorm?: {
     readonly alpha: number | readonly number[];
     readonly beta: number | readonly number[];
   };
-  /** Fill value for the recognizer canvas padding. Defaults to 128 (neutral gray). */
+  /** Recognizer canvas padding fill value. Defaults to 128 (neutral gray). */
   readonly recognizerPadValue?: number;
   /**
-   * Custom recognizer decode, replacing the built-in greedy CTC. Receives the
-   * raw `recognize_<W>` output tensor (shape `[1, T, V]`, softmaxed per the
-   * contract) and the charset, and returns the recognized text plus a confidence
-   * in `[0,1]`. Use for non-CTC heads (attention/AR decoders) or custom scoring.
-   * MUST be a worklet — it runs on the pipeline's worklet thread.
+   * Custom recognizer decode, replacing the built-in greedy CTC. Receives the raw
+   * `recognize_<W>` output tensor (shape `[1, T, V]`) and the charset, and returns the
+   * recognized text plus a confidence in `[0,1]`. Use for non-CTC heads (attention/AR
+   * decoders) or custom scoring. MUST be a worklet.
    */
   readonly decode?: (
     logits: Tensor,
@@ -106,11 +90,11 @@ export type OCROptions = {
 };
 
 /**
- * Per-run OCR options (passed to `runOCR`, not baked into the model — toggling
+ * Per-run OCR options (passed to `runOcr`, not baked into the model — toggling
  * them needs no reload).
  * @category Types
  */
-export type RunOCROptions = {
+export type RunOcrOptions = {
   /**
    * Add handling for upright stacked columns (e.g. vertical signage, shipping-
    * container codes — letters stacked top-to-bottom) on top of the normal
@@ -128,40 +112,27 @@ export type RunOCROptions = {
    * after this run, so memory doesn't accumulate as image/box sizes vary across
    * runs (worse on CoreML, which compiles a graph per method). Default `true`.
    * The document orchestrator passes `false` for its per-region OCR calls and
-   * frees once per page via `releaseMethods` instead, so it keeps the run's
+   * frees once per page via `releaseMethodsWorklet` instead, so it keeps the run's
    * working set cached while still bounding memory.
    */
   readonly release?: boolean;
 };
 
-// Defaults for the shared baked contract — the detector input is raw RGB /255
-// (mean/std baked into the PTE), the recognizer is RGB with (x/255−0.5)/0.5 norm
-// and constant-128 left padding, both heads emit softmaxed probabilities, and
-// confidence is the mean of per-character max-probs. CRAFT/DBNet decode the
-// heatmap natively; everything else can be overridden per model via OCROptions
-// (recognizerNorm/recognizerPadValue/decode, and 'custom' detectorKind+extractBoxes).
-// Per-architecture default drop score:
-const DEFAULT_DROP_SCORE: Record<'craft' | 'dbnet' | 'custom', number> = {
-  craft: 0,
-  dbnet: 0.5,
-  custom: 0,
-};
-
 /**
  * Model configuration required to instantiate an OCR task runner. One fused PTE
  * exposing `detect` + `recognize`.
  * @category Types
  */
-export type OCRModel = {
+export type OcrModel = {
   readonly modelPath: string;
-  readonly ocrOpts: OCROptions;
+  readonly ocrOpts: OcrOptions;
 };
 
 /**
  * A single recognized text region.
  * @category Types
  */
-export type OCRDetection = {
+export type OcrDetection = {
   readonly text: string;
   readonly confidence: number;
   /**
@@ -169,60 +140,40 @@ export type OCRDetection = {
    * axis-aligned box with `boundingBoxOf(quad)` from `cv.ops.boxes` if needed.
    */
   readonly quad: readonly Point[];
-  /** Wall-clock time spent recognizing this box (ms), incl. any retries. */
-  readonly recognizeMs: number;
 };
 
 /**
  * The result of one OCR run: the recognized text regions.
  * @category Types
  */
-export type OCRResult = {
-  readonly detections: OCRDetection[];
+export type OcrResult = {
+  readonly detections: OcrDetection[];
 };
 
-// Default recognizer normalization / pad (SVTR/CRNN); overridable per model via
-// OCROptions.recognizerNorm / recognizerPadValue. Detector-side norm and the
-// box-extraction tuning live with the engine in ocrPipeline.ts.
 const RECOGNIZER_ALPHA = 1 / 127.5; // (x/255 - 0.5)/0.5 -> [-1, 1]
 const RECOGNIZER_BETA = -1;
 const RECOGNIZER_PAD_VALUE = 128; // neutral gray
-// A box taller than this ratio is read as an upright stacked column.
 const TALL_CROP_RATIO = 1.5;
-// Per-page cap on stacked-column re-detection passes (each is detector-scale).
 const MAX_VERTICAL_REDETECTIONS = 8;
 // Vertical reads are lower-confidence and opt-in, so they skip the drop-score gate.
 const VERTICAL_DROP_SCORE = 0;
+// TEMP: stacked-column re-detection is disabled to measure whether char-level
+// column reading affects quality. Set true to restore it.
+const STACKED_COLUMNS_ENABLED = false;
 
-// Appends a detection when it has text and clears the drop-score threshold. A
-// module-level worklet (not a closure) so the run loop stays flat.
 function pushDetection(
-  out: OCRDetection[],
+  out: OcrDetection[],
   threshold: number,
   text: string,
   conf: number,
-  quad: readonly Point[],
-  ms: number
+  quad: readonly Point[]
 ): void {
   'worklet';
   if (text.length > 0 && conf >= threshold) {
-    out.push({ text, confidence: conf, quad, recognizeMs: ms });
+    out.push({ text, confidence: conf, quad });
   }
 }
 
-// Reorders recognized detections into human reading order (the detector emits
-// boxes in an arbitrary order). Column-aware: genuine multi-column pages read
-// column-by-column, single-column pages line-by-line, words within a line
-// left-to-right. Defined before its caller so the worklet plugin captures it.
-function orderDetections(dets: OCRDetection[]): OCRDetection[] {
-  'worklet';
-  if (dets.length <= 1) {
-    return dets;
-  }
-  const order = readingOrderIndices(dets.map((d) => d.quad));
-  return order.map((i) => dets[i]!);
-}
-
 /**
  * Creates a unified OCR runner for two-stage detect -> recognize models
  * (EasyOCR / PaddleOCR). It loads one fused PTE, validates the `detect` and
@@ -235,24 +186,20 @@ function orderDetections(dets: OCRDetection[]): OCRDetection[] {
  * @returns A promise resolving to an object with recognition and disposal
  * controls.
  */
-export async function createOCR(
-  config: OCRModel,
+export async function createOcr(
+  config: OcrModel,
   runtime?: WorkletRuntime
 ): Promise<{
   dispose: () => void;
-  runOCR: (input: ImageBuffer, options?: RunOCROptions) => Promise<OCRResult>;
-  runOCRWorklet: (input: ImageBuffer, options?: RunOCROptions) => OCRResult;
-  /** Free all bucket-method arenas without disposing the model (see `RunOCROptions.release`). */
-  releaseMethods: () => Promise<void>;
-  /** Worklet-thread variant of {@link releaseMethods}. */
+  runOcr: (input: ImageBuffer, options?: RunOcrOptions) => Promise<OcrResult>;
+  runOcrWorklet: (input: ImageBuffer, options?: RunOcrOptions) => OcrResult;
+  /** Free all bucket-method arenas without disposing the model (see `RunOcrOptions.release`). */
   releaseMethodsWorklet: () => void;
 }> {
   const { modelPath, ocrOpts } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
-  const dropScore = ocrOpts.dropScore ?? DEFAULT_DROP_SCORE[ocrOpts.detectorKind];
-  // Recognizer normalization / pad / decode — defaults preserve the SVTR/CRNN
-  // contract; OCROptions can override per model (see RecContext).
+  const dropScore = ocrOpts.dropScore ?? 0;
   const recNormAlpha = ocrOpts.recognizerNorm?.alpha ?? RECOGNIZER_ALPHA;
   const recNormBeta = ocrOpts.recognizerNorm?.beta ?? RECOGNIZER_BETA;
   const recPadValue = ocrOpts.recognizerPadValue ?? RECOGNIZER_PAD_VALUE;
@@ -260,15 +207,16 @@ export async function createOCR(
 
   const detBuckets = ocrOpts.buckets.detect;
   const recBuckets = ocrOpts.buckets.recognize;
-  // Validation + scratch allocation can throw (bad buckets, missing methods,
-  // shape/charset mismatch); on any failure dispose the model and any tensors
-  // already built, so a bad config doesn't leak native memory.
-  let recSets: RecSet[] = [];
+  // Validation + scratch allocation can throw; each tensor is pushed into
+  // `allocated` the moment it exists (one call per statement) so the catch can
+  // dispose every native allocation — a bad config must not leak.
+  const allocated: Tensor[] = [];
+  const recSets: RecSet[] = [];
   let recC = 3;
   let recH = 0;
   let charset: string[] = [];
   let recSetByWidth: ReadonlyMap<number, RecSet> = new Map();
-  let detSets: DetSet[] = [];
+  const detSets: DetSet[] = [];
   let detSetByS: ReadonlyMap<number, DetSet> = new Map();
   try {
     if (detBuckets.length === 0 || recBuckets.length === 0) {
@@ -276,30 +224,22 @@ export async function createOCR(
         'OCR: buckets.detect and buckets.recognize must each list at least one size.'
       );
     }
-    const detExtraChannels = validateDetectorSchema(
-      model,
-      detBuckets,
-      ocrOpts.detectorKind,
-      ocrOpts.extractBoxes
-    );
-
-    // Derive shapes/contract, then allocate + own the scratch tensors here so
-    // ownership never crosses a function boundary (the derive* helpers return no
-    // tensors).
-    const rec = deriveRecognizerContract(model, recBuckets);
+    const detContract = resolveDetectorContract(model, detBuckets);
+    const rec = resolveRecognizerContract(model, recBuckets);
     recC = rec.recC;
     recH = rec.recH;
-    // Push into the pre-declared arrays as we allocate, so a mid-loop tensor()
-    // failure leaves the partial set visible to the catch's dispose* below.
     for (const bucket of rec.buckets) {
-      recSets.push({
-        width: bucket.width,
-        tCanvas: tensor('uint8', [rec.recH, bucket.width, rec.recC]),
-        tCF: tensor('uint8', [rec.recC, rec.recH, bucket.width]),
-        tNorm: tensor('float32', [rec.recC, rec.recH, bucket.width]),
-        tInput: tensor('float32', bucket.inShape),
-        tLogits: tensor('float32', bucket.outShape),
-      });
+      const tCanvas = tensor('uint8', [rec.recH, bucket.width, rec.recC]);
+      allocated.push(tCanvas);
+      const tCF = tensor('uint8', [rec.recC, rec.recH, bucket.width]);
+      allocated.push(tCF);
+      const tNorm = tensor('float32', [rec.recC, rec.recH, bucket.width]);
+      allocated.push(tNorm);
+      const tInput = tensor('float32', bucket.inShape);
+      allocated.push(tInput);
+      const tLogits = tensor('float32', bucket.outShape);
+      allocated.push(tLogits);
+      recSets.push({ width: bucket.width, tCanvas, tCF, tNorm, tInput, tLogits });
     }
     recSetByWidth = new Map(recSets.map((recSet) => [recSet.width, recSet]));
 
@@ -312,42 +252,41 @@ export async function createOCR(
         `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${rec.vocabSize}).`
       );
     }
-    for (const { s, outputs } of deriveDetectorOutputs(
-      model,
-      detBuckets,
-      ocrOpts.detectorKind,
-      detExtraChannels
-    )) {
-      detSets.push({
-        s,
-        tColor: tensor('uint8', [s, s, 3]),
-        tCF: tensor('uint8', [3, s, s]),
-        tNorm: tensor('float32', [3, s, s]),
-        tInput: tensor('float32', [1, 3, s, s]),
-        tOutputs: outputs.map((o) => tensor(o.dtype, o.shape)),
-      });
+    for (const { s, outputs } of detContract) {
+      const tColor = tensor('uint8', [s, s, 3]);
+      allocated.push(tColor);
+      const tCF = tensor('uint8', [3, s, s]);
+      allocated.push(tCF);
+      const tNorm = tensor('float32', [3, s, s]);
+      allocated.push(tNorm);
+      const tInput = tensor('float32', [1, 3, s, s]);
+      allocated.push(tInput);
+      const tOutputs: Tensor[] = [];
+      for (const spec of outputs) {
+        const tOut = tensor(spec.dtype, spec.shape);
+        allocated.push(tOut);
+        tOutputs.push(tOut);
+      }
+      detSets.push({ s, tColor, tCF, tNorm, tInput, tOutputs });
     }
     detSetByS = new Map(detSets.map((detSet) => [detSet.s, detSet]));
   } catch (e) {
-    disposeRecSets(recSets);
-    disposeDetSets(detSets);
+    for (const t of allocated) {
+      t.dispose();
+    }
     model.dispose();
     throw e;
   }
 
-  // Built-in box-decode config (custom archs decode in TS via extractBoxes).
-  const extractOpts: DetectorExtractConfig | undefined = buildExtractOpts(ocrOpts.detectorKind);
-
   const dispose = () => {
     disposeRecSets(recSets);
     disposeDetSets(detSets);
     model.dispose();
   };
 
-  // Free every per-size method's activation arena (detect_<S>/recognize_<W>)
-  // without disposing the model — they transparently reload on next execute.
-  // Defined before runOCRWorklet so the worklet plugin captures it (referenced
-  // worklets must precede their callers in source order).
+  // Frees each bucket method's activation arena without disposing the model; a
+  // freed method transparently reloads on its next execute. Must precede
+  // runOcrWorklet: the worklet plugin resolves referenced worklets by source order.
   const releaseMethodsWorklet = () => {
     'worklet';
     for (const s of detBuckets) {
@@ -358,7 +297,7 @@ export async function createOCR(
     }
   };
 
-  const runOCRWorklet = (input: ImageBuffer, options?: RunOCROptions): OCRResult => {
+  const runOcrWorklet = (input: ImageBuffer, options?: RunOcrOptions): OcrResult => {
     'worklet';
     const vertical = options?.vertical ?? false;
     const tallCropRatio = options?.tallCropRatio ?? TALL_CROP_RATIO;
@@ -366,33 +305,32 @@ export async function createOCR(
     const release = options?.release ?? true;
     const { data, width, height, format } = input;
     const numChannels = FORMAT_CHANNELS[format];
-    // Both detector and recognizer read RGB, so one conversion code serves both.
-    const rgbCode = toRgbCode(format);
+    const rgbCode = FORMAT_CONVERSION[format].rgb;
 
-    // Detector state, reused for the page pass and the per-box character pass.
     const detCtx: DetectContext = {
       model,
       detBuckets,
       numChannels,
       detCode: rgbCode,
-      extractOpts,
       extractBoxes: ocrOpts.extractBoxes,
       detSets: detSetByS,
     };
 
     const tInputRaw = tensor('uint8', [height, width, numChannels]);
-    const tRecImage = tensor('uint8', [height, width, recC]);
+    let tRecImage: Tensor | null = null;
     try {
       tInputRaw.setData(data);
 
-      // ---- detector pass: letterbox -> detect_<S> -> text-box quads (image space) ----
       const quads = detectQuads(detCtx, tInputRaw, width, height);
       if (quads.length === 0) {
         return { detections: [] };
       }
 
-      // ---- recognizer source: full-res image in RGB ----
-      const recSrc = rgbCode !== null ? cvtColor(tInputRaw, tRecImage, rgbCode) : tInputRaw;
+      let recSrc = tInputRaw;
+      if (rgbCode !== null) {
+        tRecImage = tensor('uint8', [height, width, recC]);
+        recSrc = cvtColor(tInputRaw, tRecImage, rgbCode);
+      }
       const recCtx: RecContext = {
         model,
         recSetByWidth,
@@ -404,20 +342,16 @@ export async function createOCR(
         padValue: recPadValue,
         decode: recDecode,
       };
-      // The vertical path crops each box from the raw page and re-detects its
-      // characters; `recCode`/`recC` convert a box crop to RGB.
       const vctx: VerticalContext = {
         detCtx,
         rawPage: tInputRaw,
-        recCode: rgbCode,
         recC,
         tallCropRatio,
         redetectBudget: { remaining: maxRedetections },
       };
 
-      const detections: OCRDetection[] = [];
+      const detections: OcrDetection[] = [];
 
-      // Valid (non-tiny) boxes, ordered TL,TR,BR,BL.
       const ordered: Point[][] = [];
       for (const quad of quads) {
         const orderedQuad = orderQuad(quad);
@@ -429,21 +363,14 @@ export async function createOCR(
 
       if (!vertical) {
         for (const orderedQuad of ordered) {
-          const boxStart = nowMs();
           const { text, conf } = recognizeQuad(recCtx, recSrc, orderedQuad);
-          pushDetection(detections, dropScore, text, conf, orderedQuad, nowMs() - boxStart);
+          pushDetection(detections, dropScore, text, conf, orderedQuad);
         }
-        return { detections: orderDetections(detections) };
+        return { detections: orderByReadingOrder(detections) };
       }
 
-      // Additive vertical pass: read x-aligned stacked glyph boxes as one joined
-      // column word; everything else (lines, isolated boxes) reads normally.
       const { columns, singles } = groupVerticalColumns(ordered);
       for (const col of columns) {
-        const boxStart = nowMs();
-        // `recognizeGlyphStrip` splits any multi-letter box into single-glyph
-        // cells (DBNet merges stacked letters and won't split them), so the
-        // column's boxes can be passed straight through, top -> bottom.
         const strip = recognizeGlyphStrip(recCtx, recSrc, col);
         if (strip) {
           pushDetection(
@@ -451,44 +378,33 @@ export async function createOCR(
             VERTICAL_DROP_SCORE,
             strip.text,
             strip.conf,
-            boundingQuadOf(col),
-            nowMs() - boxStart
+            boundingQuadOf(col)
           );
         }
       }
       for (const orderedQuad of singles) {
-        const boxStart = nowMs();
-        const { text, conf, stacked } = readBoxVertical(
-          recCtx,
-          vctx,
-          recSrc,
-          orderedQuad,
-          quadSize(orderedQuad)
-        );
-        pushDetection(
-          detections,
-          stacked ? VERTICAL_DROP_SCORE : dropScore,
-          text,
-          conf,
-          orderedQuad,
-          nowMs() - boxStart
-        );
+        const size = quadSize(orderedQuad);
+        if (STACKED_COLUMNS_ENABLED && size.height >= size.width * vctx.tallCropRatio) {
+          const stacked = readStackedColumn(recCtx, vctx, orderedQuad, size);
+          if (stacked) {
+            pushDetection(detections, VERTICAL_DROP_SCORE, stacked.text, stacked.conf, orderedQuad);
+            continue;
+          }
+        }
+        const { text, conf } = recognizeQuad(recCtx, recSrc, orderedQuad);
+        pushDetection(detections, dropScore, text, conf, orderedQuad);
       }
-      return { detections: orderDetections(detections) };
+      return { detections: orderByReadingOrder(detections) };
     } finally {
       tInputRaw.dispose();
-      tRecImage.dispose();
-      // Standalone runs free their bucket arenas so memory stays bounded as
-      // sizes vary; the document orchestrator opts out (release: false) and
-      // frees once per page.
+      tRecImage?.dispose();
       if (release) {
         releaseMethodsWorklet();
       }
     }
   };
 
-  const runOCR = wrapAsync(runOCRWorklet, runtime);
-  const releaseMethods = wrapAsync(releaseMethodsWorklet, runtime);
+  const runOcr = wrapAsync(runOcrWorklet, runtime);
 
-  return { runOCR, runOCRWorklet, dispose, releaseMethods, releaseMethodsWorklet };
+  return { runOcr, runOcrWorklet, dispose, releaseMethodsWorklet };
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrSymbols.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/charsets.ts
similarity index 99%
rename from packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrSymbols.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/charsets.ts
index 64c7c40c14..f8a151299e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrSymbols.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/charsets.ts
@@ -155,7 +155,7 @@ export const symbols = {
  * Supported OCR language codes (EasyOCR alphabets).
  * @category Types
  */
-export type OCRLanguage = keyof typeof symbols;
+export type OcrLanguage = keyof typeof symbols;
 
 export const PPOCR_SYMBOLS =
   '!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz¢£¤¥¦§¨©ª«¬®¯°±²³´µ¶·¸¹º»¼½¾¿ÀÁÂÃÄÅÆÇÈÉÊËÌÍÎÏÐÑÒÓÔÕÖ×ØÙÚÛÜÝÞßàáâãäåæçèéêëìíîïðñòóôõö÷øùúûüýþÿĀāĂ㥹ĆćĈĉĊċČčĎďĐđĒēĔĕĖėĘęĚěĜĝĞğĠġĢģĤĥĦħĨĩĪīĬĭĮįİıIJijĴĵĶķĸĹĺĻļĽľĿŀŁłŃńŅņŇňʼnŊŋŌōŎŏŐőŒœŔŕŖŗŘřŚśŜŝŞşŠšŢţŤťŦŧŨũŪūŬŭŮůŰűŲųŴŵŶŷŸŹźŻżŽžſƀƁƂƃƄƅƆƇƈƉƊƋƌƍƎƏƐƑƒƓƔƕƖƗƘƙƚƛƜƝƞƟƠơƢƣƤƥƦƧƨƩƪƫƬƭƮƯưƱƲƳƴƵƶƷƸƹƺƻƼƽƾƿǀǁǂǃDŽDždžLJLjljNJNjnjǍǎǏǐǑǒǓǔǕǖǗǘǙǚǛǜǝǞǟǠǡǢǣǤǥǦǧǨǩǪǫǬǭǮǯǰDZDzdzǴǵǶǷǸǹǺǻǼǽǾǿȀȁȂȃȄȅȆȇȈȉȊȋȌȍȎȏȐȑȒȓȔȕȖȗȘșȚțȜȝȞȟȠȡȢȣȤȥȦȧȨȩȪȫȬȭȮȯȰȱȲȳȴȵȶȷȸȹȺȻȼȽȾȿɀɁɂɃɄɅɆɇɈɉɊɋɌɍɎɏɐɑɒɓɔɕɖɗɘəɚɛɜɝɞɟɠɡɢɣɤɥɦɧɨɩɪɫɬɭɮɯɰɱɲɳɴɵɶɷɸɹɺɻɼɽɾɿʀʁʂʃʄʅʆʇʈʉʊʋʌʍʎʏʐʑʒʓʔʕʖʗʘʙʚʛʜʝʞʟʠʡʢʣʤʥʦʧʨʩʪʫʬʭʮʯΑΒΓΔΕΖΗΘΙΚΛΜΝΞΟΠΡΣΤΥΦΧΨΩαβγδεζηθικλμνξοπρστυφχψωϐϑϒϕϖϝϞϟϠϡϢϣϤϥϦϧϨϩϪϫϬϭϮϯϰϱϴϵẠồ–—―‖‘’“”†‡•‥…‰′″※⁎⁰⁴⁵⁶⁷⁸⁹⁺⁻⁼⁽⁾ⁿ₀₁₂₃₄₅₆₇₈₉₊₋₌₍₎ₒ₠₡₢₣₤₥₦₧₨₩₪₫€₭₮₯₰₱₲₳₴₵₶₷₸₹₺₻₼₽₾₿℃℉ℎℏℑ℘ℜ™℧Åℵℶℷℸℹ⅀ⅠⅡⅢⅣⅤⅥⅦⅧⅨⅩⅪⅫⅰⅱⅲⅳⅴⅵⅶⅷⅸⅹⅺⅻⅼↀↁↂ←↑→↓↔↕↖↗↘↙↚↛↜↝↞↟↠↡↢↣↤↥↦↧↨↩↪↫↬↭↮↯↰↱↲↳↴↵↶↷↸↹↺↻↼↽↾↿⇀⇁⇂⇃⇄⇅⇆⇇⇈⇉⇊⇋⇌⇍⇎⇏⇐⇑⇒⇓⇔⇕⇖⇗⇘⇙⇚⇛⇜⇝⇞⇟⇠⇡⇢⇣⇤⇥⇦⇧⇨⇩⇪⇫⇬⇭⇮⇯⇰⇱⇲⇳⇴⇵⇶⇷⇸⇹⇺⇻⇼⇽⇾⇿∀∁∂∃∄∅∆∇∈∉∋∏∑−∓∕∖∙√∛∜∝∞∟∠∡∢∥∧∨∩∪∫∬∭∮∯∰∱∲∳∴∵∶∷∸∹∺∻∼∽∾∿≀≁≂≃≄≅≆≇≈≉≊≋≌≍≎≏≐≑≒≓≔≕≖≗≘≙≚≛≜≝≞≟≠≡≢≣≤≥≦≧≨≩≪≫≬≭≮≯≰≱≲≳≴≵≶≷≸≹≺≻≼≽≾≿⊀⊁⊂⊃⊄⊅⊆⊇⊈⊉⊊⊋⊌⊍⊎⊏⊐⊑⊒⊓⊔⊕⊖⊗⊘⊙⊚⊛⊜⊝⊞⊟⊠⊡⊢⊣⊤⊥⊦⊧⊨⊩⊪⊫⊬⊭⊮⊯⊰⊱⊲⊳⊴⊵⊶⊷⊸⊹⊺⊻⊼⊽⊾⊿⋅⌀⌃⌘⌚⌛⌤⌥⌦⌧⌨〉⌫⌬⌭⌮⌯⍵⍺⎆⎇⎈⎉⎊⎋⎌⎍⎎⎏⎐⎑⎒⎓⎔⎕⎖⎗⎘⎙⎚⎛⎜⎝⎞⎟⎠⎡⎢⎣⎤⎥⎦⎧⎨⎩⎪⎫⎬⎭⎮⎯⎰⎱⎲⎳⎴⎵⎶⎷⎸⎹⎺⎻⎼⎽⎾⎿⏀⏁⏂⏃⏄⏅⏆⏇⏈⏉⏊⏋⏌⏍⏎⏏⏐⏑⏒⏓⏔⏕⏖⏗⏘⏙⏚⏛⏜⏝⏞⏟⏠⏡⏢⏣⏤⏥⏦⏧⏨⏩⏪⏫⏬⏭⏮⏯⏰⏱⏲⏳⏴⏵⏶⏷⏸⏹⏺⏻⏼⏽⏾⏿①②③④⑤⑥⑦⑧⑨⑩⑪⑫⑬⑭⑮⑯⑰⑱⑲⑳Ⓜ━╆╳▁█■□▪▫▬▲△▶▷▼▽◀◆◇◉◊○◎●◐◥◻◼◽◾☀☁☂☃☄★☆☇☈☉☊☋☌☍☎☏☐☑☒☓☔☕☖☗☘☙☚☛☜☝☞☟☠☡☢☣☤☥☦☧☨☩☪☫☬☭☮☯☰☱☲☳☴☵☶☷☸☹☺☻☼☽☾☿♀♁♂♃♄♅♆♇♈♉♊♋♌♍♎♏♐♑♒♓♔♕♖♗♘♙♚♛♜♝♞♟♠♡♢♣♤♥♦♧♨♩♪♫♬♭♮♯♰♱♲♳♴♵♶♷♸♹♺♻♼♽♾♿⚀⚁⚂⚃⚄⚅⚆⚇⚈⚉⚊⚋⚌⚍⚎⚏⚐⚑⚒⚓⚔⚕⚖⚗⚘⚙⚚⚛⚜⚝⚞⚟⚠⚡⚢⚣⚤⚥⚦⚧⚨⚩⚪⚫⚬⚭⚮⚯⚰⚱⚲⚳⚴⚵⚶⚷⚸⚹⚺⚻⚼⚽⚾⚿⛀⛁⛂⛃⛄⛅⛆⛇⛈⛉⛊⛋⛌⛍⛎⛏⛐⛑⛒⛓⛔⛕⛖⛗⛘⛙⛚⛛⛜⛝⛞⛠⛡⛢⛣⛤⛥⛦⛧⛨⛩⛪⛫⛬⛭⛮⛯⛲⛳⛵⛶⛺⛽⛾⛿✂✅✆✇✈✉✊✋✌✍✎✏✐✑✒✓✔✕✖✘✙✚✛✜✝✞✟✠✡✢✣✤✥✦✧✨✩✪✫✬✭✮✯✰✱✲✳✴✵✶✷✸✹✺✻✼✽✾✿❀❁❂❃❄❅❆❇❈❉❊❋❌❍❎❏❐❑❒❓❔❕❖❗❘❙❚❛❜❝❞❡❢❣❤❥❦❧❨❩❪❫❬❭❮❯❰❱❲❳❴❵❶❷❸❹❺❻❼❽❾❿➀➁➂➃➄➅➆➇➈➉➊➋➌➍➎➏➐➑➒➓➔➕➖➗➘➙➚➛➜➝➞➟➠➡➢➣➤➥➦➧➨➩➪➫➬➭➮➯➰➱➲➳➴➵➶➷➸➹➺➻➼➽➾➿⟶⟹⤴⤵⬅⬆⬇⬛⬜⭐⭕　、。々〆〇〈〉《》「」『』【】〒〔〕〖〗〜〰〵〽ぁあぃいぅうぇえぉおかがきぎくぐけげこごさざしじすずせぜそぞただちぢっつづてでとどなにぬねのはばぱひびぴふぶぷへべぺほぼぽまみむめもゃやゅゆょよらりるれろゎわゑをんゕゖゝゞァアィイゥウェエォオカガキギクグケゲコゴサザシジスズセゼソゾタダチヂッツヅテデトドナニヌネノハバパヒビピフブプヘベペホボポマミムメモャヤュユョヨラリルレロヮワヰヲンヴヵヶヷヸヹヺ・ーヽヾヿ㈱㊗㊙㐂㐱㑇㑊㑳㑺㒺㕁㕑㕥㕮㘎㘭㙍㙘㙟㙦㚷㛃㛚㛹㝛㝠㝡㟁㟃㠇㠓㠣㠯㠶㡌㢘㤘㤙㥄㥫㥮㧐㧑㧟㧱㨗㨪㩗㩦㩳㪚㪟㫰㬉㬊㬎㬚㬪㭎㭕㮣㮾㰀㳄㳇㳒㳘㳚㴔㴪㴱㵐㶲㸃㸆㸌㺄㻬㼝㽏㽞㿠䁖䂮䃅䃎䃮䅟䈰䊀䌹䎃䎖䏝䏡䏲䐃䓖䓛䓣䓨䓫䓬䖝䗖䗛䗪䗬䗴䘏䘑䘚䜣䝉䝔䝙䠀䠶䢺䢼䣘䥽䦃䮄䰟䰾䲁䲟䲠䲢䴉䴓䴔䴕䴖䴗䴘䴙䶊䶮一丁七丄万丈三上下丌不与丏丐丑专且丕世丘丙业丛东丝丞丟両丢两严並丧个丫中丮丰丱串临丸丹为主丼丽举丿乂乃久么义之乌乍乎乏乐乒乓乔乖乗乘乙乜九乞也习乡书乩乭买乱乳乸乹乾亀亂了予争亊事二亍于亏云互亓五井亘亙亚些亜亞亟亡亢交亥亦产亨亩享京亭亮亯亰亱亲亳亵亶亷亸亹人亾亿什仁仂仃仄仅仆仇仉今介仍从仏仑仓仔仕他仗付仙仛仜仝仞仟仡代令以仨仩仪仫们仮仰仱仲仳仵件价仺任仼份仿伀企伃伈伉伊伋伍伎伏伐休伔伕众优伙会伛伝伞伟传伢伣伤伥伦伧伪伫伬伭伯估伱伲伴伶伷伸伺伻似伽伾佀佁佃但佇佈佉佋佌位低住佐佑佒体佔何佖佗佘余佚佛作佝佞佟你佢佣佤佥佧佩佪佬佯佰佳佴併佶佷佸佹佺佻佼佽佾使侀侁侂侃侄侅來侇侈侉侊例侍侏侐侑侔侖侗侘侚供侜依侞侠価侣侥侦侧侨侩侪侬侮侯侲侳侴侵侶侷侹侻便俁係促俄俅俇俉俊俋俌俍俎俏俐俑俓俔俗俘俙俚俛俜保俞俟俠信俣俤俦俨俩俪俫俬俭修俯俱俳俴俵俶俷俸俺俽俾倀倃倅倆倇倈倉個倌倍倏倐們倒倓倔倕倖倗倘候倚倛倜倞借倠倡倢倣値倥倦倧倨倩倪倫倬倭倮倰倳倴倶倷倸倹债倻值倾偁偃偅假偈偉偊偋偌偍偎偏偐偓偕偘做偛停偝偞偟偠偡偢健偨偩偪偫偬偭偯偰偲側偵偶偷偸偺偻偽偾偿傀傂傃傅傈傉傋傌傍傎傑傒傔傕傖傘備傚傛傜傝傞傢傣傥傧储傩催傭傮傯傰傱傲傳傴債傷傺傻傽傾傿僁僂僄僅僆僇僈僉僊僋働僎像僑僓僔僕僖僗僙僚僛僜僝僞僡僤僥僦僧僨僩僪僬僭僮僰僱僳僴僵僶僸價僻僽僾僿儀儁儂儃億儆儇儈儉儋儌儐儑儒儓儔儕儗儘儚儛儜儞償儠儡儢儤儥儦優儫儰儱儲儳儴儵儷儸儹儺儻儼儽儿兀允元兄充兆兇先光克兌免兎児兑兒兔兕兖兗兙党兛兜兝兞兠兡兢兣入內全兩兪八公六兮兰共关兴兵其具典兹养兼兽兿冀冁冂冄内円冇冈冉冊册再冏冐冑冒冓冔冕冗冘写冚军农冞冠冢冣冤冥冧冨冪冬冮冯冰冱冲决冴况冶冷冺冻冼冽净凃凄凅准凇凈凉凊凋凌凍减凑凔凖凘凛凜凝凞几凡凢凤処凪凫凬凭凯凰凱凳凵凶凸凹出击凼函凿刀刁刂刃刄分切刈刉刊刌刍刎刑划刓刖列刘则刚创刜初刞删判別刦刧刨利刪别刬刭刮到刱刲刳刵制刷券刹刺刻刼刽刿剀剁剂剃剄剅剆則剉削剋剌前剎剏剐剑剒剔剕剖剗剚剛剜剝剞剟剡剣剤剥剧剩剪剫剬剭剮副剰割剳剴創剷剸剹剺剻剼剽剿劀劁劂劃劄劇劈劉劊劋劌劍劐劑劓劔劖劗劘劙力劝办功加务劢劣劦动助努劫劬劭励劲劳労劵効劼劾势勁勃勅勇勉勋勌勍勐勑勒勔動勖勗勘務勚勛勝勞募勢勣勤勦勧勩勰勱勲勳勴勵勷勸勺勻勾勿匀匁匂匃匄包匆匈匊匋匍匏匐匕化北匙匜匝匟匠匡匢匣匦匪匭匮匯匰匱匲匴匷匸匹区医匼匽匾匿區十千卅升午卉半卋卌卍华协卐卑卒卓協单卖南単博卜卞卟占卡卢卣卤卦卧卫卬卮卯印危卲即却卵卷卸卹卺卻卼卽卿厀厂厄厅历厉压厌厍厎厒厓厔厕厖厗厘厙厚厜厝厞原厠厢厣厤厥厦厨厩厬厭厮厰厲厳厴厹去厾县叁参參叄叆叇又叉及友双反収发叔叕取受变叙叚叛叟叠叡叢口古句另叨叩只叫召叭叮可台叱史右叵叶号司叹叻叼叽吁吃各吆合吉吊吋同名后吏吐向吒吓吔吕吖吗吚君吝吞吟吠吡吣吥否吧吨吩吪含听吭吮启吰吱吲吳吴吵吶吷吸吹吻吼吽吾呀呂呃呆呇呈呉告呋呌呎呐呑呒呓呔呕呖呗员呙呛呜呢呣呤呥呦周呪呫呬呰呱呲味呴呵呶呷呸呺呻呼命呾呿咀咁咂咄咅咆咇咈咉咊咋和咍咎咏咐咑咒咔咕咖咗咘咙咚咛咝咟咠咡咢咣咤咥咦咧咨咩咪咫咬咭咮咯咱咲咳咴咶咷咸咺咻咼咽咾咿哀品哂哃哄哅哆哇哈哉哌响哎哏哐哑哒哓哔哕哖哗哙哚哝哞哟員哢哤哥哦哧哨哩哪哫哭哮哱哲哳哶哺哼哽哿唁唃唄唅唆唇唈唉唌唏唐唑唒唔唕唖唗唘唚唛唝唠唢唣唤唦唧唪唫唬唭售唯唰唱唲唳唴唵唶唷唸唻唼唾唿啁啃啄啅商啉啊啍啎問啐啑啒啓啕啖啗啚啜啞啟啡啢啣啤啥啦啧啪啫啬啭啮啯啰啱啲啴啵啶啷啸啻啼啽啾喀喁喂喃善喆喇喈喉喊喋喌喍喏喑喒喓喔喘喙喚喜喝喟喢喣喤喥喦喧喨喩喪喫喬喭單喰喱喲喳喵営喷喹喻喼喽喾喿嗀嗁嗃嗄嗅嗆嗇嗈嗉嗊嗋嗌嗍嗎嗏嗐嗑嗒嗓嗔嗕嗖嗚嗛嗜嗝嗞嗟嗡嗢嗣嗤嗥嗦嗧嗨嗩嗪嗫嗬嗮嗯嗲嗳嗵嗶嗷嗹嗺嗼嗽嗾嗿嘀嘁嘂嘄嘅嘆嘈嘉嘌嘍嘎嘏嘐嘑嘒嘓嘔嘕嘖嘗嘘嘚嘛嘜嘝嘞嘟嘠嘡嘢嘣嘤嘥嘧嘩嘬嘭嘮嘯嘰嘱嘲嘳嘴嘵嘶嘷嘸嘹嘻嘽嘾嘿噀噁噂噅噆噇噈噉噊噌噍噎噏噓噔噗噘噙噚噛噜噝噞噠噢噣噤噥噦器噩噪噫噬噭噮噯噰噱噲噳噴噶噸噹噺噻噼嚀嚃嚄嚅嚆嚇嚈嚌嚍嚎嚏嚐嚒嚓嚕嚗嚘嚙嚚嚜嚞嚟嚢嚣嚤嚥嚦嚧嚨嚩嚪嚫嚬嚭嚮嚯嚲嚳嚴嚵嚶嚷嚼嚽嚾嚿囀囁囂囃囅囆囈囉囊囋囌囍囑囒囓囔囗囙囚四囝回囟因囡团団囤囧囪囫园囮囯困囱囲図围囵囶囷囹固国图囿圀圁圂圃圄圅圆圇圈圉圊國圌圍圏圐園圓圖圗團圙圜圞土圠圢圣圧在圩圪圫圬圭圮圯地圲圳圹场圻圾址坁坂坅均坉坊坋坌坍坎坏坐坑坒坖块坚坛坜坝坞坟坠坡坣坤坥坦坨坩坪坫坬坭坮坯坰坱坲坳坴坵坶坷坻坼坽坿垀垂垃垄垆垈型垌垍垎垏垒垓垔垕垗垘垙垚垛垜垝垞垟垠垡垢垣垤垦垧垩垫垭垮垯垰垱垲垴垵垸垺垻垼垽垾垿埂埃埅埆埇埈埋埌城埏埒埔埕埗埘埙埚埜埝域埠埡埢埤埧埨埪埫埭埮埯埳埴埵埶執埸培基埻埼埽堀堂堃堅堆堇堈堉堊堋堌堍堎堐堑堔堕堖堘堙堝堞堠堡堣堤堥堦堧堨堩堪堮堯堰報堲場堵堷堺堼堽堾堿塀塁塄塅塆塈塉塊塋塌塍塏塑塒塓塔塕塗塘塙塚塜塝塞塟塡塢塤塥塨塩填塬塭塯塰塱塲塴塵塸塹塺塼塽塾塿墀墁境墅墆墈墉墊墋墎墐墒墓墕墖増墘墙墚墜墝增墟墠墡墣墦墨墩墪墫墬墮墯墱墳墶墺墻墼墽墾壁壄壅壆壇壈壊壋壌壎壐壑壓壔壕壖壘壙壚壛壝壞壟壠壢壤壩壪士壬壮壯声壱売壳壴壶壷壸壹壺壻壼壽壾壿夀夃处夆备変夊夋夌复夎夏夐夒夔夕外夗夘夙多夜够夠夢夤夥大天太夫夬夭央夯失头夷夸夹夺夼夾奀奁奂奄奅奇奈奉奊奋奎奏奐契奓奔奕奖套奘奚奠奡奢奣奥奧奨奩奪奫奬奭奮奰奲女奴奶奸她奼好妀妁如妃妄妆妇妈妊妍妎妏妐妑妒妓妖妗妘妙妝妞妠妢妣妤妥妦妧妨妩妪妫妬妭妮妯妲妳妵妷妸妹妺妻妾姀姁姃姆姈姉姊始姌姍姎姏姐姑姒姓委姗姘姙姚姜姝姞姡姣姤姥姦姨姪姫姬姮姱姵姶姸姹姺姻姼姽姿娀威娃娄娅娆娇娈娉娋娌娍娎娑娓娖娘娙娛娜娟娠娣娥娩娫娭娯娱娲娳娴娵娶娷娸娼娽娿婀婁婆婉婊婌婍婐婑婓婕婗婘婚婞婟婠婢婣婤婥婦婧婪婫婬婭婯婳婴婵婶婷婸婺婻婼婿媂媃媄媆媊媐媒媓媔媕媖媗媚媛媜媞媟媢媥媦媧媪媬媭媮媯媰媱媲媳媵媸媺媻媼媽媾媿嫁嫂嫄嫇嫈嫉嫋嫌嫐嫒嫔嫕嫖嫗嫘嫚嫛嫜嫠嫡嫢嫣嫥嫦嫩嫪嫫嫭嫮嫰嫱嫲嫳嫴嫵嫶嫷嫺嫻嫽嫿嬀嬃嬅嬈嬉嬋嬌嬎嬐嬓嬖嬗嬙嬛嬝嬡嬢嬤嬥嬨嬪嬬嬭嬰嬲嬴嬷嬸嬽嬾嬿孀孁孃孅孆孈孋孌子孑孓孔孕孖字存孙孚孛孜孝孟孢季孤孥学孩孪孫孬孮孰孱孲孳孵孷學孺孻孼孽孿宁宂它宄宅宇守安宋完宍宎宏宓宕宗官宙定宛宜宝实実宠审客宣室宥宦宧宨宪宫宬宮宰害宴宵家宸容宽宾宿寀寁寂寃寄寅密寇寈寊寋富寍寐寑寒寓寔寕寖寗寘寙寛寜寝寞察寠寡寢寤寥實寧寨審寪寫寬寮寯寰寱寳寵寶寸对寺寻导対寿封専尃射尅将將專尉尊尋尌對導小尐少尒尓尔尕尖尗尘尙尚尜尝尟尢尤尥尧尨尪尬尭尰就尳尴尷尸尹尺尻尼尽尾尿局屁层屃屄居屆屇屈屉届屋屌屍屎屏屐屑屓屔展屖屘屙屚屜屝属屠屡屢屣層履屦屧屨屩屬屭屮屯山屴屹屺屻屼屾屿岀岁岂岅岈岉岊岋岌岍岏岐岑岒岔岖岗岘岙岚岛岜岝岞岟岠岡岢岣岤岧岨岩岪岫岬岭岮岯岰岱岳岵岷岸岻岽岿峁峂峃峄峅峇峉峊峋峌峍峐峒峗峘峙峚峛峝峞峠峡峣峤峥峦峧峨峩峪峬峭峮峯峰峱峴島峷峻峼峽峿崀崁崂崃崄崆崇崈崋崌崍崎崏崐崑崒崔崖崗崘崙崚崛崝崞崟崠崡崢崣崤崥崦崧崨崩崬崭崮崰崱崳崴崵崶崷崺崽崾崿嵁嵂嵃嵅嵇嵊嵋嵌嵎嵐嵑嵒嵕嵖嵗嵘嵙嵚嵛嵜嵝嵞嵣嵥嵨嵩嵫嵬嵮嵯嵰嵱嵲嵴嵷嵺嵻嵽嵾嵿嶀嶁嶂嶃嶄嶅嶆嶇嶈嶉嶊嶋嶌嶍嶒嶓嶔嶕嶗嶙嶚嶜嶝嶞嶟嶠嶡嶢嶦嶧嶨嶩嶪嶬嶭嶮嶰嶱嶲嶴嶵嶷嶸嶺嶼嶽巀巂巃巄巅巆巇巉巋巌巍巎巏巑巒巔巕巖巗巘巛川州巟巡巢巣工左巧巨巩巫差巯巰己已巳巴巵巶巷巹巻巽巾巿帀币市布帅帆师帊希帏帐帑帔帕帖帗帘帙帚帛帜帝帟帠帡帢帣帤帥带帧帨帩師帬席帮帯帰帱帳帴帶帷常帻帼帽帾幀幂幃幄幅幇幌幎幏幑幓幔幕幖幗幘幙幚幛幝幞幟幠幡幢幣幦幧幨幩幪幫幬幭幰干平年幵并幷幸幹幺幻幼幽幾广庀庁庂広庄庆庇庈庉床庋庌序庐庑库应底庖店庙庚庛府庞废庠庢庣庤庥度座庨庪庫庬庭庮庰庱庲庳庴庵庶康庸庹庻庼庽庾廁廂廃廄廅廆廇廈廉廊廋廌廍廎廐廑廒廓廔廕廖廗廘廙廚廛廜廝廞廟廠廡廢廣廥廦廧廨廩廪廬廮廯廰廱廳廵延廷廸廹建廻廼廾廿开弁异弃弄弅弆弇弈弉弊弋弌弍式弐弑弒弓弔引弖弗弘弚弛弝弟张弢弣弤弥弦弧弨弩弭弮弯弱弳張弶強弸弹强弼弾彀彃彄彅彆彈彉彊彋彌彎彏彐归当彔录彖彗彘彙彜彝彞彟彠彡形彣彤彥彦彧彩彪彫彬彭彯彰影彳彴彷彸役彻彼彾彿往征徂徃径待徇很徉徊律後徐徑徒従徕得徘徙徛徜從徠御徥徦徧徨復循徫徬徭微徯徲徳徴徵德徸徹徼徽徾徿忀忁心必忆忉忌忍忏忐忑忒忔忕忖志忘忙応忝忞忠忡忣忤忥忧忨忪快忬忭忮忯忱忳念忷忸忺忻忼忽忾忿怀态怂怃怄怅怆怊怋怍怎怏怐怒怓怔怕怖怗怙怚怛怜思怞怠怡怢急怦性怨怩怪怫怬怭怮怯怱怲怳怴怵怷怹总怼怿恀恁恂恃恅恆恇恉恊恋恌恍恐恒恓恔恕恙恚恛恝恞恟恠恡恢恣恤恥恧恨恩恪恫恬恭息恰恲恳恵恶恸恹恺恻恼恽恿悀悁悃悄悅悆悇悈悉悊悌悍悐悒悔悕悖悗悚悛悝悞悟悠悢患悤悦悧您悩悪悫悬悭悯悰悱悲悳悴悵悶悷悸悹悺悻悼悽悾惄情惆惇惈惉惊惋惌惎惏惑惓惔惕惘惙惚惛惜惝惟惠惡惢惣惤惥惦惧惨惩惪惫惬惭惮惯惰惱惲想惴惵惶惷惸惹惺惻惼惾愀愁愃愅愆愈愉愊愍愎意愐愒愓愔愕愖愚愛愜感愠愣愤愦愧愨愫愬愭愮愯愲愴愶愷愼愽愾愿慁慂慄慅慆慇慈慉慊態慌慍慎慑慒慓慔慕慘慚慛慜慝慞慟慢慣慤慥慧慨慪慫慬慭慮慰慱慲慳慴慵慶慷慹慺慼慽慾憀憂憃憇憉憊憋憍憎憐憑憒憓憔憕憖憙憚憛憝憟憡憢憤憧憨憩憪憫憬憭憮憯憰憱憲憴憵憶憷憸憺憻憼憾憿懁懂懃懅懆懇懈應懊懋懌懍懐懑懒懔懘懞懟懠懣懤懥懦懧懨懪懫懭懮懰懱懲懵懶懷懸懺懻懼懽懾懿戀戁戃戄戆戇戈戉戊戋戌戍戎戏成我戒戔戕或戗战戙戚戛戞戟戠戡戢戣戤戥戦戧戩截戬戭戮戯戰戱戲戳戴戶户戸戹戺戻戼戽戾房所扁扂扃扅扆扇扈扉扊手才扎扐扑扒打扔払托扙扛扜扞扡扢扣扤扥扦执扩扪扫扬扭扮扯扰扱扳扶批扺扻扼扽找承技抃抄抆抇抉把抌抎抏抑抒抓抔投抖抗折抚抛抜抝択抟抠抡抢护报抦抨抪披抬抭抮抱抴抵抶抹抻押抽抾抿拂拃拄担拆拇拈拉拊拋拌拍拎拏拐拑拒拓拔拕拖拗拘拙拚招拜拝拟拠拡拢拣拤拥拦拧拨择拫括拭拮拯拱拲拳拴拵拶拷拸拹拺拼拽拾拿挀持挂挃指挈按挋挌挍挎挏挐挑挓挖挙挚挛挝挞挟挠挡挣挤挥挦挨挩挪挫挬振挱挲挴挵挶挹挺挻挼挽挾挿捁捂捃捄捅捆捉捊捋捌捍捎捏捐捒捔捕捖捗捘捜捞损捡换捣捥捦捧捨捩捫捭据捯捱捲捵捶捷捺捻捼捽掀掂掃掄掇授掉掊掌掎掏掐排掔掖掗掘掙掛掜掝掞掟掠採探掣掤接控推掩措掫掬掭掮掯掰掱掲掳掴掷掸掺掻掼掽掾揀揃揄揅揆揉揊揍揎描提揑插揔揕揖揘揚換揜揝揟揠握揣揤揥揨揩揪揫揭揮揯揱揲揳援揵揶揷揸揹揺揽揿搀搁搂搅搆搉搊搋搌損搏搐搒搓搔搕搖搗搘搚搛搜搞搟搠搡搢搣搤搥搦搧搨搪搫搬搭搯搰搳搴搵搶搷搹携搽搾摀摁摂摃摄摅摆摇摈摊摋摍摎摏摐摑摒摓摔摘摛摜摝摞摟摠摡摦摧摩摭摮摯摰摲摳摴摵摶摸摹摺摻摽摿撂撃撄撅撇撈撊撋撌撏撐撑撒撓撕撖撗撙撚撜撝撞撟撠撢撣撤撥撦撩撫撬播撮撰撲撳撵撷撸撹撺撻撼撽撾撿擀擁擂擄擅擇擉擊擋操擎擏擐擒擓擔擕擖擗擘據擞擠擡擢擣擤擥擦擧擨擩擫擬擭擯擰擱擲擴擷擸擺擻擼擽擾擿攀攄攆攇攉攌攍攏攐攒攓攔攕攖攗攘攙攛攜攝攞攠攡攢攣攤攥攦攩攪攫攬攭攮支攲攳攴收攷攸改攻攽放政敁敂故敆效敉敊敌敍敎敏救敓敔敕敖敗敘教敛敜敝敞敟敢散敤敦敧敩敪敫敬敭敯数敲敳整敵敶敷數敹敺敻敾敿斀斁斂斃斄文斉斋斌斎斐斑斒斓斔斕斖斗料斛斜斝斞斟斠斡斤斥斧斨斩斪斫斬断斮斯新斲斵斶斷方斻於施斿旁旂旃旄旅旆旉旋旌旍旎族旐旒旓旖旗旙旚旛旝旞旟无旡既日旦旧旨早旬旭旮旯旰旱旲旳旴旵时旷旸旺旻旼旽旾旿昀昂昃昄昆昇昈昉昊昌昍明昏昐昒易昔昕昙昚昛昜昝昞星映昡昢昣昤春昦昧昨昪昫昬昭是昰昱昲昳昴昵昶昺昼昽显晁時晃晅晇晈晉晊晋晌晏晐晒晓晔晕晖晗晙晚晛晜晝晞晟晡晢晤晥晦晧晨晩晪晫晬晭普景晰晱晳晴晶晷晸智晻晼晾暀暁暂暃暄暅暆暇暈暉暊暋暌暍暎暏暐暑暕暖暗暘暝暟暠暡暢暦暧暨暩暫暮暰暱暲暴暵暶暸暹暻暾暿曀曄曅曆曇曈曉曊曋曌曏曒曔曖曙曚曛曜曝曠曡曣曦曧曨曩曬曭曮曰曲曳更曶曷書曹曺曼曽曾替最朁會朄朅月有朊朋服朏朐朒朓朔朕朖朗朘望朝朞期朡朢朣朦朧木未末本札朮术朱朳朴朵朶朸朹机朻朼朽朾朿杀杁杂权杄杅杆杇杈杉杋杌李杏材村杓杕杖杗杙杜杝杞束杠条杢杣杤来杧杨杩杪杬杭杮杯杰東杲杳杴杵杶杷杻杼松板极构枅枆枇枉枋枌枍枎枏析枒枓枕枖林枘枚枛果枝枞枠枡枢枣枥枧枨枪枫枭枯枰枱枲枳枴枵架枷枸枹枺枻枼柀柁柂柃柄柅柈柉柊柍柎柏某柑柒染柔柖柘柙柚柜柝柞柟柠柢柣柤查柧柩柫柬柮柯柰柱柲柳柴柵柶柷柸柹柺査柽柾柿栀栁栂栃栄栅标栈栉栊栋栌栎栏栐树栒栓栔栖栗栘栚栜栝栞栟校栢栨栩株栫栭栯栰栱栲栳栴栵样核根栻格栽栾桀桁桂桃桄桅框案桉桊桋桌桎桏桐桑桒桓桔桕桖桙桜桝桟桠桡桢档桤桥桦桧桨桩桫桭桮桯桱桲桴桵桶桷桹桺桼桾桿梀梁梂梃梅梆梇梉梌梏梐梒梓梔梗梘梛梜條梟梠梡梢梣梦梧梨梩梫梬梭梮梯械梱梲梳梴梵梶梼梽梾梿检棁棂棃棄棆棉棊棋棌棍棐棒棓棕棖棗棘棚棛棜棟棠棡棣棤棧棨棩棪棫棬森棯棰棱棲棳棵棶棷棸棹棺棻棼棽椀椁椄椅椆椇椈椉椋椌植椎椏椐椑椒椓椔椗椙検椟椠椤椥椪椭椰椲椳椴椵椶椸椹椽椿楀楂楄楅楈楉楊楎楒楓楔楕楖楗楘楙楚楛楝楞楟楠楡楢楣楤楥楦楧楨楩楪楫楬業楮楯楰楱楳極楶楷楸楹楺楼楽楿概榃榄榅榆榇榈榉榊榍榎榑榔榕榖榗榘榙榛榜榞榠榢榣榤榦榧榨榩榪榫榬榭榮榯榰榱榲榴榶榷榹榻榼榽榾榿槀槁槃槄槇槉槊構槌槍槎槐槑槓槔槕槗様槙槚槛槜槟槠槢槤槥槧槨槩槬槭槮槰槱槲槳槷槸槺槻槼槽槾槿樀樁樂樅樆樊樋樑樓樔樕樗樘標樛樝樞樟樠模樣樧樨権横樫樯樱樲樴樵樸樹樺樻樽樾樿橀橁橄橇橈橉橋橐橑橒橓橘橙橚橛橜橝橞機橡橢橤橥橦橧橨橪橫橭橱橶橹橼橾橿檀檁檃檄檇檉檊檍檎檐檑檓檔檕檖檗檛檜檝檞檟檠檡檢檣檥檦檨檩檫檬檮檯檳檵檸檹檻檽櫂櫃櫅櫆櫈櫋櫌櫐櫑櫓櫙櫚櫛櫜櫝櫞櫟櫠櫡櫥櫧櫨櫪櫫櫬櫰櫱櫳櫸櫹櫺櫻櫼櫾櫿欂欃欄欅欈欉權欋欏欐欑欒欖欘欞欠次欢欣欤欥欧欨欬欭欯欱欲欳欴欵欶欷欸欹欺欻欽款欿歂歃歅歆歇歈歉歊歋歌歍歎歐歓歔歕歙歛歜歟歠歡止正此步武歧歩歪歭歯歲歳歴歶歷歸歹死歼歾歿殀殁殂殃殄殆殇殈殉殊残殍殏殑殒殓殔殖殗殘殙殚殛殞殟殠殡殢殣殤殥殦殧殪殫殭殮殯殰殲殳殴段殶殷殺殻殼殽殿毀毁毂毃毄毅毆毇毈毉毊毋毌母毎每毐毒毓比毕毖毗毘毙毚毛毞毡毣毤毦毨毪毫毬毯毰毲毳毴毵毸毹毻毼毽毾毿氀氁氂氄氅氆氇氈氊氌氍氏氐民氓气氕氖気氘氙氚氛氟氠氡氢氣氤氦氧氨氩氪氫氬氮氯氰氲氳水氶氷永氹氻氽氾氿汀汁求汃汆汇汈汉汊汋汍汎汏汐汒汔汕汗汙汚汛汜汝汞江池污汤汦汧汨汩汪汫汭汯汰汱汲汳汴汶汸汹決汽汾沁沂沃沄沅沆沇沈沉沋沌沍沏沐沒沓沔沕沖沘沙沚沛沜沟没沢沣沤沥沦沧沨沩沪沫沬沭沮沯沱河沴沶沷沸油沺治沼沽沾沿況泂泃泄泅泆泇泉泊泌泐泒泓泔法泖泗泙泚泛泜泝泞泠泡波泣泥泧注泩泪泫泬泭泮泯泰泱泲泳泵泷泸泺泻泼泽泾洀洁洄洇洈洊洋洌洍洎洑洒洓洗洘洙洚洛洞洟洢洣洤津洧洨洩洪洫洭洮洱洲洳洴洵洶洸洹洺活洼洽派洿流浃浄浅浆浇浈浉浊测浍济浏浐浑浒浓浔浕浘浙浚浛浜浞浟浠浡浢浣浤浥浦浧浩浪浬浭浮浯浰浲浴浵浶海浸浹浺浼浿涂涄涅涆涇消涉涊涌涍涎涐涑涒涓涔涕涖涗涘涙涛涝涞涟涠涡涢涣涤润涧涨涩涪涫涬涮涯液涳涴涵涷涸涼涽涾涿淀淄淅淆淇淈淊淋淌淍淎淏淑淒淓淔淖淘淙淚淛淜淝淞淟淠淡淢淤淥淦淨淩淪淫淬淭淮淯淰深淳淴淵淶混淸淹淺添淼淽渀渃清渇済渉渊渋渌渍渎渐渑渓渔渕渗渙渚減渜渝渟渠渡渢渣渤渥渦渨温渫測渭渮港渰渱渲渳渴游渹渺渻渼渽渾渿湀湁湃湄湅湆湇湉湊湋湍湎湑湓湔湖湘湛湜湝湞湟湠湡湢湣湥湧湨湩湫湮湯湱湲湳湴湻湼湾湿満溁溃溅溆溇溈溉溋溍溎溏源溓溔準溘溙溚溛溜溝溞溟溠溢溥溦溧溪溫溮溯溰溱溲溳溴溵溶溷溹溺溻溼溽溾滀滁滂滃滄滅滆滇滈滉滋滌滍滎滏滑滒滓滔滕滖滗滘滙滚滛滜滝滞滟滠满滢滤滥滦滧滨滩滪滫滬滭滮滯滱滲滴滵滷滸滹滻滽滾滿漁漂漃漅漆漇漈漉漊漋漎漏漑漒漓演漕漖漘漙漚漠漢漣漤漥漦漧漩漪漫漬漭漮漯漰漱漲漳漴漵漶漷漸漹漻漼漾漿潀潁潃潄潅潆潇潋潍潎潏潐潑潒潔潕潖潗潘潙潚潛潜潝潞潟潠潢潤潦潩潬潭潮潯潰潲潳潴潵潶潷潸潺潼潽潾潿澀澁澂澄澅澆澇澈澉澋澌澍澎澐澒澓澔澖澗澛澜澞澠澡澣澤澥澦澧澨澩澪澫澭澮澯澱澳澴澶澹澻澼澽激濁濂濃濄濆濇濈濉濊濋濌濎濑濒濔濕濘濙濛濜濞濟濠濡濣濤濦濧濩濫濬濭濮濯濰濱濲濴濶濺濻濼濾濿瀀瀁瀄瀅瀆瀇瀉瀋瀌瀍瀎瀏瀑瀔瀕瀖瀘瀙瀚瀛瀝瀞瀟瀠瀡瀢瀣瀦瀧瀨瀩瀫瀬瀯瀰瀱瀲瀴瀵瀶瀷瀸瀹瀺瀼瀾瀿灂灃灄灈灉灊灋灌灏灑灕灖灗灘灛灝灞灟灡灣灤灧灨灩灪火灭灯灰灴灵灶灸灼災灾灿炀炁炅炆炉炊炌炎炒炔炕炖炘炙炜炝炟炣炤炫炬炭炮炯炰炱炲炳炷炸点為炻炼炽炾烀烁烂烃烈烉烊烋烏烑烒烓烔烖烘烙烛烜烝烟烠烤烦烧烨烩烫烬热烯烰烱烴烶烷烹烺烻烽焃焄焆焉焊焌焍焐焓焔焕焖焗焘焙焚焜焞焠無焣焦焮焯焰焱焴然焻焼焿煁煃煅煆煇煉煊煋煌煎煐煑煒煓煔煕煖煗煙煚煜煝煞煟煠煢煣煤煥煦照煨煩煬煮煲煳煴煵煶煸煺煽熀熂熄熅熇熈熉熊熏熐熒熔熖熗熘熙熚熛熜熟熠熤熥熨熬熯熰熱熲熳熵熸熹熺熻熼熾熿燀燁燂燃燄燅燈燉燊燋燎燏燐燒燔燕燖燗燘燙燚燜燝營燠燡燥燦燧燫燬燭燮燴燹燻燼燾燿爀爁爂爆爇爊爌爍爐爓爔爕爚爛爝爞爟爢爣爨爪爬爭爯爰爱爲爵父爷爸爹爺爻爼爽爾爿牀牁牂牄牆片版牉牋牌牍牎牏牒牓牕牖牘牙牚牛牝牟牠牡牢牣牤牥牦牧物牬牮牯牰牲牴牵牷牸特牺牻牼牽牾牿犀犁犂犄犅犆犇犈犉犊犋犌犍犎犏犐犑犒犓犕犖犗犘犚犛犝犞犟犠犢犣犤犥犦犧犨犩犪犬犮犯犰犴犵状犷犸犹犽犿狀狁狂狃狄狆狈狉狊狋狌狍狎狐狑狒狓狔狖狗狘狙狚狛狜狝狞狟狠狡狣狤狥狦狨狩狪狫独狭狮狯狰狱狲狳狴狶狷狸狹狺狻狼狽狾狿猀猁猂猃猄猇猈猊猋猎猏猑猒猓猕猖猗猘猙猛猜猝猞猟猡猢猣猥猦猧猨猩猪猫猬猭献猯猰猱猲猳猴猵猶猷猹猺猻猼猾猿獀獁獂獃獄獅獇獊獋獌獍獎獏獐獑獒獗獘獛獝獞獟獠獡獢獣獥獦獧獨獩獪獫獬獭獮獯獰獲獳獴獵獶獷獸獺獻獼獽獾獿玀玁玂玃玄玅玆率玈玉玊王玎玏玑玒玓玔玕玖玗玘玙玚玛玞玟玠玡玢玤玥玦玧玨玩玫玭玮环现玱玲玳玶玷玹玺玻玼玾玿珀珂珅珇珈珉珊珋珌珍珎珏珐珑珒珓珕珖珙珛珝珞珠珡珢珣珤珥珦珧珩珪珫班珮珰珲珵珶珷珸珹珺珽現琀琁球琄琅理琇琈琉琊琍琎琏琐琔琖琚琛琟琡琢琤琥琦琨琪琫琬琭琮琯琰琱琲琳琴琵琶琹琺琼琿瑀瑁瑂瑃瑄瑅瑆瑇瑈瑊瑋瑑瑒瑓瑔瑕瑖瑗瑙瑚瑛瑜瑝瑞瑟瑠瑢瑣瑤瑥瑧瑨瑩瑪瑬瑭瑯瑰瑱瑲瑳瑴瑵瑶瑷瑺瑽瑾瑿璀璁璃璄璅璆璇璈璉璊璋璌璎璐璒璕璗璘璙璚璜璞璟璠璡璢璣璥璦璧璨璩璪璫璬璮環璱璲璵璸璹璺璽璿瓀瓁瓅瓈瓊瓋瓌瓏瓑瓒瓔瓖瓘瓚瓛瓜瓝瓞瓟瓠瓡瓢瓣瓤瓥瓦瓨瓩瓬瓮瓯瓴瓵瓶瓷瓻瓽瓾瓿甀甂甃甄甈甋甌甍甏甐甑甒甓甔甕甖甗甘甙甚甜甝甞生甡產産甥甦用甩甪甫甬甭甮甯田由甲申电男甸甹町画甽甾甿畀畅畇畈畊畋界畎畏畐畑畔畖留畚畛畜畝畟畠畢畣畤略畦畧番畫畬畯異畱畲畳畴畵當畷畸畹畺畽畿疀疁疃疆疇疊疋疌疍疎疏疐疑疒疔疕疖疗疘疙疚疝疟疠疡疢疣疤疥疧疫疬疭疮疯疰疱疲疳疴疵疸疹疻疼疽疾疿痀痁痂痃痄病症痈痉痊痋痌痍痎痏痐痑痒痓痔痕痗痘痙痚痛痝痞痟痠痡痢痣痤痦痧痨痩痪痫痯痰痱痲痳痴痷痸痹痺痻痼痾痿瘀瘁瘃瘅瘆瘈瘉瘊瘋瘌瘍瘏瘐瘑瘓瘕瘖瘗瘘瘙瘛瘜瘝瘞瘟瘠瘡瘢瘣瘤瘥瘦瘧瘩瘪瘫瘭瘮瘯瘰瘱瘲瘳瘴瘵瘸瘺瘻瘼瘽瘾瘿癀療癃癄癅癆癇癈癉癌癍癐癒癓癔癖癗癘癙癜癞癟癠癡癢癣癤癥癧癩癪癫癬癭癮癯癰癱癲癵癸癹発登發白百癿皁皂的皆皇皈皊皋皎皏皐皑皓皕皖皙皚皛皜皝皞皤皦皫皭皮皯皰皱皲皴皵皷皸皺皻皽皾皿盂盃盄盅盆盇盈盉益盋盌盍盎盏盐监盒盓盔盖盗盘盛盜盝盞盟盡監盤盥盦盧盨盩盪盫盬盭目盯盱盲盳直盷相盹盺盻盼盾眀省眃眄眅眇眈眉眊看県眍眐眑眒眓眕眙眚眛眜眝眞真眠眢眣眥眦眨眩眬眭眯眱眲眳眴眵眶眷眸眹眺眼眽眾着睁睃睄睅睆睇睊睋睌睍睎睏睐睑睒睔睕睖睚睛睜睞睟睠睡睢督睥睦睧睨睩睪睫睬睭睮睯睹睺睼睽睾睿瞀瞁瞂瞄瞅瞇瞉瞋瞌瞍瞎瞏瞑瞒瞓瞖瞘瞚瞛瞜瞝瞞瞟瞠瞡瞢瞥瞧瞩瞪瞫瞬瞭瞰瞲瞳瞴瞵瞶瞷瞻瞼瞽瞿矁矂矇矉矊矌矍矎矏矐矓矔矕矗矘矙矚矛矜矞矠矢矣知矧矩矫矬短矮矯矰矱矲石矴矶矸矹矺矻矼矽矾矿砀码砂砄砅砆砉砋砌砍砎砏砐砑砒研砕砖砗砘砚砜砝砟砠砢砣砥砦砧砨砩砫砬砭砮砯砰砲砳破砵砷砸砹砺砻砼砾础硁硂硃硅硇硈硉硊硌硍硎硏硐硒硓硔硕硖硗硙硚硜硝硠硤硥硨硩硪硫硬硭确硯硰硱硼硾硿碁碃碄碅碆碇碈碉碌碍碎碏碑碓碔碕碗碘碚碛碜碞碟碡碣碥碧碨碩碪碫碭碰碱碲碳碴碶碸碹確碻碼碽碾磁磃磅磈磉磊磋磌磍磎磏磐磑磔磕磘磙磚磛磜磝磞磟磡磣磥磧磨磩磪磬磯磱磲磳磴磵磷磹磺磻磼磽磾磿礁礂礄礅礉礌礎礐礑礒礓礔礙礛礜礝礞礠礣礥礦礧礨礪礫礬礭礮礱礴礵示礼礽社礿祀祁祂祃祅祆祇祈祉祊祋祌祎祏祐祑祒祓祔祕祖祗祙祚祛祜祝神祟祠祡祢祣祤祥祧票祩祪祫祭祯祲祳祴祷祸祹祺祼祾祿禀禁禂禃禄禅禇禊禋禍禎福禑禒禓禔禕禖禗禘禚禛禜禟禠禡禤禦禧禨禩禪禫禬禭禮禰禱禳禴禵禷禸禹禺离禼禽禾禿秀私秃秅秆秈秉秊秋种秎秏科秒秔秕秖秘租秠秣秤秦秧秩秪秫秬秭秮积称秳秶秸秺移秽秾稀稂稃稅稆稈稉稊程稌稍税稑稒稔稗稘稙稚稜稞稟稠稣稫稬稭種稯稰稱稲稳稷稹稺稻稼稽稾稿穀穂穄穅穆穇穈穉穊穋穌積穎穏穑穖穗穙穛穜穟穠穡穢穣穤穧穨穩穫穬穭穮穰穱穴穵究穷穸穹空穽穾穿窀突窃窄窅窆窈窉窊窋窌窍窎窏窐窑窒窓窔窕窖窗窘窙窜窝窞窟窠窢窣窥窦窨窩窪窫窬窭窮窯窰窱窲窳窴窵窶窸窺窻窾窿竀竁竃竄竅竇竈竊立竑竖竘站竚竜竝竞竟章竢竣童竦竩竪竫竭端竴競竷竹竺竻竽竿笃笄笅笆笈笊笋笏笐笑笓笔笕笘笙笛笝笞笠笢笤笥符笨笩笪笫第笭笮笯笰笱笲笳笴笵笸笹笺笻笼笾筀筄筅筆筇筈等筊筋筌筍筏筐筑筒答策筘筚筛筜筝筞筠筡筢筤筥筦筧筩筬筭筮筯筰筱筲筳筴筵筶筷筹筻筼签简箄箅箆箇箈箊箋箌箍箎箏箐箑箒箓箔箕箖算箘箙箛箜箝箠管箢箦箧箨箩箪箫箬箭箯箱箴箷箸箹箾節篁範篆篇築篊篋篌篎篑篓篔篕篙篚篛篝篞篠篡篣篤篥篦篧篨篩篪篭篮篯篱篲篳篴篷篸篹篻篼篽篾篿簀簁簂簃簅簇簉簋簌簍簎簏簑簒簕簖簗簙簜簝簞簟簠簡簢簣簥簦簧簨簩簪簫簬簭簮簰簳簷簸簹簺簻簼簽簾簿籀籁籃籈籉籊籌籍籐籔籗籙籛籜籝籟籠籣籤籥籦籧籩籪籬籮籯籲米籴籸籹籺类籼籽籾粂粃粄粇粉粊粋粍粑粒粔粕粗粘粛粜粝粞粟粢粤粥粦粧粩粪粮粱粲粳粵粹粺粻粼粽精粿糀糁糅糇糈糉糊糌糍糎糒糔糕糖糗糙糜糝糞糟糠糢糧糨糪糬糯糰糱糲糴糵糶糷糸糹糺系糾紀紂紃約紅紆紇紈紉紊紋納紎紐紑紒紓純紕紖紗紘紙級紛紜紝紞紟素紡索紥紧紩紫紬紮累細紱紲紳紵紶紸紹紺紼紽紾紿絀絁終絃組絅絆絇経絎絏結絓絕絖絘絛絜絝絞絡絢絣給絧絨絪絫絭絮絯絰統絲絳絵絶絷絹絺絻絼絽絿綀綁綃綄綅綆綈綉綌綍綎綏綑經綔綖継続綜綝綞綟綠綢綣綦綧綪綫綬維綮綯綰綱網綳綴綵綷綸綹綺綻綼綽綾綿緀緁緂緃緄緅緆緇緈緊緋緌緎総緐緑緒緔緖緗緘緙線緛緜緝緞締緡緣緤緥緦緧編緩緪緬緯緰緱緲緳練緵緶緷緹緺緻縁縂縃縄縅縈縉縊縋縌縎縏縐縑縒縓縕縗縚縛縜縝縞縟縠縡縢縣縤縦縧縩縪縫縭縮縯縰縱縲縳縴縵縶縷縸縹縻縼總績縿繀繁繂繃繄繅繆繇繈繉繊繋繍繐繑繒織繕繖繘繙繚繜繝繞繟繠繡繢繣繦繧繩繪繫繭繮繯繰繲繳繴繵繶繷繸繹繺繻繼繽繾纀纁纂纆纇纈纊纋續纍纏纐纑纒纓纔纕纖纗纘纚纛纜纠纡红纣纤纥约级纨纩纪纫纬纭纮纯纰纱纲纳纴纵纶纷纸纹纺纻纼纽纾线绀绁绂练组绅细织终绉绊绋绌绍绎经绐绑绒结绔绕绖绗绘给绚绛络绝绞统绠绡绢绣绤绥绦继绨绩绪绫续绮绯绰绱绲绳维绵绶绷绸绹绺绻综绽绾绿缀缁缂缃缄缅缆缇缈缉缊缌缎缐缑缒缓缔缕编缗缘缙缚缛缜缝缞缟缠缡缢缣缤缥缦缧缨缩缪缫缬缭缮缯缰缱缲缳缴缵缶缸缹缺缽缾缿罂罃罄罅罇罈罉罊罋罌罍罎罏罐网罔罕罗罘罚罛罜罝罞罟罠罡罢罣罥罦罧罨罩罪罫罬罭置罰罱署罳罴罵罶罷罹罺罻罼罽罾罿羁羂羃羅羆羇羈羉羊羋羌羍美羑羒羓羔羕羖羗羙羚羛羜羝羞羟羠羡羢羣群羥羦羧羨義羬羭羯羰羱羲羳羴羵羶羷羸羹羺羼羽羾羿翀翁翂翃翄翅翇翈翊翋翌翍翎翏翐翑習翔翕翘翙翚翛翜翟翠翡翢翣翥翦翨翩翪翫翬翭翮翯翰翱翲翳翴翷翹翺翻翼翽翾翿耀老考耄者耆耇耋而耍耎耏耐耑耒耔耕耖耗耘耙耜耞耟耠耡耢耤耥耦耧耨耩耪耬耮耰耱耳耴耵耶耷耸耹耻耽耾耿聂聃聆聇聈聊聋职聍聏聑聒联聖聘聚聝聞聟聡聨聩聪聯聰聱聲聳聴聵聶職聹聽聾聿肂肃肄肅肆肇肉肋肌肎肏肐肒肓肕肖肘肙肚肛肜肝肟肠股肢肣肤肥肧肩肪肫肭肮肯肱育肴肵肷肸肹肺肼肽肾肿胀胁胂胃胄胅胆胇胈胉胊背胍胎胏胐胑胔胕胖胗胘胙胚胛胜胝胞胠胡胣胤胥胧胨胩胪胫胬胭胯胰胱胲胳胴胵胶胷胸胹胺胻胼能胾脀脁脂脃脅脆脇脈脉脊脍脎脏脐脑脒脓脔脕脖脗脘脙脚脛脞脟脡脣脤脥脧脩脫脬脭脯脰脱脲脳脶脷脸脹脽脾脿腃腄腆腇腈腊腋腌腍腎腏腐腑腒腓腔腕腖腘腙腚腛腜腞腠腢腤腥腦腧腨腩腫腭腮腯腰腱腲腳腴腶腷腸腹腺腻腼腽腾腿膀膂膃膆膇膈膉膊膋膌膍膏膑膕膘膙膚膛膜膝膞膟膠膢膣膥膦膨膩膫膬膮膰膱膲膳膴膵膷膹膺膻膽膾膿臀臂臃臄臅臆臇臈臉臊臋臌臍臏臐臑臒臓臕臗臘臙臚臛臜臝臞臟臠臡臢臣臥臦臧臨臩自臬臭臮臯臲至致臷臸臺臻臼臾臿舀舁舂舄舅舆與興舉舊舋舌舍舎舐舑舒舔舕舖舗舘舛舜舝舞舟舠舡舢舣舥舨舩航舫般舭舯舰舱舲舳舴舵舶舷舸船舺舻舼舽舾艀艂艄艅艇艉艋艎艏艐艑艒艓艔艕艘艙艚艛艟艤艦艨艫艮良艰艱色艳艴艵艶艷艸艹艺艼艽艾艿芀节芃芄芅芈芊芋芍芎芏芐芑芒芓芔芗芘芙芚芛芜芝芞芟芠芡芣芤芥芦芧芨芩芪芫芬芭芮芯芰花芲芳芴芵芷芸芹芺芻芼芽芾芿苀苁苄苅苇苈苉苊苋苌苍苎苏苑苒苓苔苕苖苗苘苙苛苜苝苞苟苠苡苣苤若苦苧苨苪苫苬苯英苲苳苴苵苶苷苹苺苻苼苾茀茁茂范茄茅茆茇茈茉茋茌茍茎茏茑茓茔茕茖茗茘茙茚茛茜茝茠茢茤茥茦茧茨茩茪茫茬茭茮茯茱茲茳茴茵茶茷茸茹茺茻茼茽茿荀荁荂荃荄荅荆荇草荊荍荎荏荐荑荒荓荔荖荘荙荚荛荜荞荟荠荡荣荤荥荦荧荨荩荪荫荬荭荮药荳荴荵荷荸荺荻荼荽荾莁莃莅莆莉莊莋莌莍莎莏莐莒莓莔莕莖莘莙莛莜莝莞莠莢莣莤莥莧莨莩莪莫莬莰莱莲莳莴莵莶获莸莹莺莼莽莿菀菁菂菅菆菇菈菉菊菋菌菍菎菏菑菓菔菕菖菘菙菜菝菞菟菠菡菣菤菥菧菨菩菪菫華菰菱菲菳菴菵菶菸菹菺菼菽菾菿萁萃萄萆萇萉萊萋萌萍萎萏萐萑萒萘萚萜萝萠萡萣萤营萦萧萨萩萬萭萯萰萱萲萳萴萵萷萸萹萺萼落萿葀葂葃葅葆葉葊葋葌葍葎葑葒葖著葙葚葛葜葝葞葟葡董葤葥葦葧葩葫葬葭葯葰葱葳葴葵葶葷葸葹葺葽蒂蒄蒆蒇蒈蒉蒋蒌蒍蒎蒏蒐蒓蒔蒗蒘蒙蒚蒛蒜蒝蒞蒟蒠蒡蒢蒤蒧蒨蒩蒪蒫蒬蒭蒮蒯蒱蒲蒴蒶蒸蒹蒺蒻蒼蒽蒾蒿蓀蓁蓂蓄蓆蓇蓉蓊蓋蓌蓍蓎蓏蓐蓑蓓蓖蓘蓚蓝蓟蓠蓢蓣蓥蓦蓧蓨蓩蓪蓫蓬蓭蓮蓯蓰蓱蓲蓳蓴蓶蓷蓹蓺蓻蓼蓽蓾蓿蔀蔂蔃蔄蔆蔇蔈蔉蔊蔌蔍蔎蔏蔑蔒蔓蔔蔕蔖蔗蔘蔚蔜蔝蔞蔟蔠蔡蔣蔤蔥蔦蔧蔨蔩蔪蔫蔬蔭蔮蔰蔱蔴蔵蔷蔸蔹蔺蔻蔼蔽蔾蔿蕀蕁蕃蕄蕅蕆蕇蕈蕉蕊蕋蕍蕎蕑蕒蕓蕕蕖蕗蕘蕙蕚蕛蕝蕞蕟蕠蕡蕢蕣蕤蕦蕧蕨蕩蕪蕫蕭蕮蕰蕱蕲蕴蕵蕷蕸蕹蕺蕻蕼蕾蕿薀薁薂薃薄薅薆薇薈薉薊薋薌薍薎薏薐薑薔薕薖薗薘薙薚薛薜薝薞薟薠薡薢薣薤薦薧薨薩薪薫薬薭薮薯薰薱薲薳薴薵薶薷薸薹薺薽薾薿藀藁藂藃藄藅藆藇藈藉藋藍藎藏藐藑藒藓藔藕藗藘藙藚藜藝藟藠藡藣藤藥藦藨藩藪藫藬藭藰藱藲藴藶藷藸藹藺藻藼藾藿蘀蘁蘂蘄蘅蘆蘇蘉蘊蘋蘌蘐蘑蘓蘖蘗蘘蘙蘚蘛蘜蘞蘟蘠蘡蘢蘣蘤蘥蘦蘧蘩蘪蘬蘭蘮蘱蘳蘴蘵蘶蘸蘹蘺蘻蘼蘾蘿虀虃虆虇虈虉虋虌虎虏虐虑虒虓虔處虖虙虚虛虜虞號虡虢虣虤虧虨虩虪虫虬虭虮虯虰虱虴虵虷虸虹虺虻虼虽虾虿蚀蚁蚂蚄蚅蚆蚇蚊蚋蚌蚍蚑蚓蚕蚖蚗蚘蚙蚚蚜蚝蚞蚡蚢蚣蚤蚥蚧蚨蚩蚪蚬蚯蚰蚱蚲蚳蚴蚵蚶蚷蚸蚹蚺蚻蚼蚿蛀蛁蛂蛃蛄蛅蛆蛇蛈蛉蛊蛋蛌蛍蛎蛏蛐蛑蛓蛔蛕蛖蛗蛘蛙蛚蛛蛜蛝蛞蛟蛢蛣蛤蛦蛩蛪蛫蛬蛭蛮蛯蛰蛱蛲蛳蛴蛵蛶蛷蛸蛹蛺蛻蛾蜀蜁蜂蜃蜄蜆蜇蜈蜉蜊蜋蜌蜍蜎蜐蜑蜒蜓蜕蜖蜗蜘蜙蜚蜛蜜蜞蜠蜡蜢蜣蜤蜥蜦蜧蜨蜩蜪蜬蜭蜮蜯蜰蜱蜲蜳蜴蜵蜷蜸蜺蜻蜼蜾蜿蝀蝁蝂蝃蝆蝇蝈蝉蝋蝌蝍蝎蝐蝑蝒蝓蝔蝕蝖蝗蝘蝙蝚蝛蝜蝝蝞蝟蝠蝡蝣蝤蝥蝦蝨蝪蝫蝬蝭蝮蝯蝰蝱蝲蝳蝴蝶蝷蝸蝹蝺蝻蝼蝽蝾蝿螁螂螃螄螅螇螈螋融螎螏螐螑螒螓螔螖螗螘螚螛螜螝螞螟螠螡螢螣螤螥螨螪螫螬螭螮螯螰螱螳螴螵螶螷螸螹螺螻螼螽螾螿蟀蟁蟂蟃蟄蟅蟆蟇蟈蟉蟊蟋蟌蟎蟏蟑蟒蟓蟔蟗蟘蟙蟛蟜蟞蟟蟠蟡蟢蟣蟤蟥蟦蟧蟨蟪蟫蟬蟭蟮蟯蟲蟳蟴蟶蟷蟹蟺蟻蟼蟾蟿蠀蠁蠂蠃蠄蠅蠆蠈蠉蠊蠋蠌蠍蠏蠐蠑蠓蠔蠕蠖蠗蠘蠙蠛蠜蠝蠟蠠蠡蠢蠣蠤蠥蠦蠨蠩蠪蠫蠬蠭蠮蠯蠰蠱蠲蠳蠵蠶蠷蠸蠹蠻蠼蠽蠾蠿血衁衂衃衄衅衆衇衈衊衋行衍衎衒術衔衕衖街衙衚衛衜衝衞衠衡衢衣补衧表衩衪衫衬衭衮衯衰衱衲衵衶衷衹衺衽衾衿袀袁袂袃袄袅袆袈袉袋袌袍袑袒袓袕袖袗袘袙袚袛袜袞袟袠袡袢袤袧袨袪被袬袭袯袰袱袲袴袵袶袷袸袹袺袼袽袾袿裀裁裂裃装裆裈裉裊裋裍裎裏裒裔裕裖裗裘裙裚裛補裝裞裟裠裡裢裣裤裥裧裨裬裮裯裰裱裲裳裴裵裶裷裸裹裺裻裼製裾裿褀褁褂褄褅褆複褊褋褌褍褎褐褑褒褓褔褕褖褗褘褙褚褛褞褟褡褢褥褦褧褩褪褫褭褮褯褰褱褲褳褴褵褶褷褸褻褼褽褾襁襂襃襄襆襇襋襌襏襐襒襓襕襖襗襘襚襛襜襝襞襟襠襡襢襣襤襦襩襪襫襬襭襮襯襲襳襴襶襷襹襺襻襾西要覂覃覅覆覇覈見覌覎規覓覕視覗覘覚覛覜覝覡覢覣覤覦覧覩親覬覭覮覯覲観覶覷覺覽覿觀见观觃规觅视觇览觉觊觋觌觎觏觐觑角觓觔觕觖觙觚觜觝觞觟觠觡觢解觤觥触觩觫觬觭觯觰觱觲觳觴觶觷觸觺觻觼觾觿言訁訂訃訄訇計訊訌討訏訐訑訒訓訔訕訖託記訚訛訝訞訟訢訣訥訧訩訪訬設訰許訳訴訶訹診註証訾訿詀詁詄詅詆詈詍詎詐詒詔評詖詗詘詙詛詝詞詟詠詡詢詣詥試詧詩詪詫詬詭詮詰話該詳詵詶詷詹詻詼詿誂誃誄誅誆誇誉誊誋誌認誑誒誓誕誖誘誙誚語誠誡誣誤誥誦誨說誫説読誰課誴誶誸誹誻誼誾調諀諂諄諅諆談諈諉請諌諍諏諐諑諒諓諔諕論諗諚諛諜諝諞諟諠諡諢諤諦諧諪諫諭諮諰諱諲諳諴諵諶諷諸諺諼諾謀謁謂謄謅謇謈謊謋謌謍謎謏謐謑謒謔謕謖謗謘謙謚講謜謝謞謠謡謢謣謤謥謦謧謨謩謪謫謬謯謰謱謳謵謷謹謼謾譀譁譅譆譇譈證譊譋譌譎譏譐譑譓譔譕譖識譙譚譛譜譝譞譟譠譣譥警譨譪譫譬譭譯議譲譳譴護譸譹譺譻譽譾譿讀讂讃讄讆變讋讌讎讐讒讓讔讕讖讘讙讚讜讞讟计订讣认讥讦讧讨让讪讫训议讯记讱讲讳讴讵讶讷许讹论讻讼讽设访诀证诂诃评诅识诇诈诉诊诋诌词诎诏诐译诒诓诔试诖诗诘诙诚诛诜话诞诟诠诡询诣诤该详诧诨诩诫诬语诮误诰诱诲诳说诵请诸诹诺读诼诽课诿谀谁谂调谄谅谆谇谈谊谋谌谍谎谏谐谑谒谓谔谕谖谗谙谚谛谜谝谞谟谠谡谢谣谤谥谦谧谨谩谪谫谬谭谮谯谰谱谲谳谴谵谶谷谹谻谼谽谾谿豁豃豅豆豇豈豉豊豋豌豍豎豏豐豔豕豗豚豜豝豟象豢豤豥豦豨豪豫豬豭豮豯豰豱豲豳豵豶豷豸豹豺豻豽豿貀貁貂貄貅貆貉貊貌貍貏貐貑貒貓貔貕貗貘貙貛貜貝貞負財貢貣貤貧貨販貪貫責貭貮貯貰貲貳貴貶買貸貺費貼貽貾貿賀賁賂賃賄賅資賈賉賊賌賎賑賒賓賔賕賙賚賛賜賞賟賠賡賢賣賤賥賦賧賨質賫賬賭賮賴賵賸賹賺賻購賽賾贄贅贆贇贈贊贋贌贍贏贐贓贔贕贖贗贙贛贜贝贞负贡财责贤败账货质贩贪贫贬购贮贯贰贱贲贳贴贵贶贷贸费贺贻贼贽贾贿赀赁赂赃资赅赆赇赈赉赊赋赌赍赎赏赐赑赒赓赔赕赖赗赘赙赚赛赜赝赞赟赠赡赢赣赤赦赧赨赩赪赫赬赭赮走赳赴赵赶起赸趀趁趂趄超越趋趌趍趎趏趐趑趓趔趕趖趙趜趟趠趡趣趥趧趨趪趫趬趭趮趯趱趲足趴趵趷趸趹趺趼趾趿跂跃跄跅跆跇跈跋跌跍跎跏跐跑跓跕跖跗跘跙跚跛跜距跞跟跠跡跢跣跤跥跦跧跨跩跪跫跬跮路跰跱跲跳跴践跶跷跸跹跺跻跼跽跾跿踂踃踅踆踇踉踊踌踍踎踏踐踑踒踓踔踕踖踘踙踛踝踞踟踠踡踢踣踤踥踦踧踩踪踫踬踮踯踰踱踳踴踵踶踸踹踺踼踽踾蹀蹁蹂蹄蹅蹇蹈蹉蹊蹋蹌蹍蹎蹏蹐蹑蹒蹓蹔蹕蹖蹗蹙蹚蹛蹜蹝蹞蹟蹠蹡蹢蹣蹤蹥蹦蹧蹩蹪蹬蹭蹯蹰蹲蹳蹴蹵蹶蹸蹺蹻蹼蹽蹾蹿躁躂躄躅躆躇躈躉躊躋躍躏躐躑躒躓躔躕躖躗躙躚躜躝躞躟躠躡躣躤躥躦躨躩躪身躬躭躯躰躲躳躶躺躽軀車軋軌軍軎軏軑軒軓軔軘軛軝軞軟転軤軥軦軧軨軫軬軮軯軱軲軵軶軷軸軹軺軻軼軽軾軿輀輁輂較輄輅輆輇輈載輊輋輌輐輑輒輓輔輕輖輗輘輚輛輜輝輞輟輠輣輤輥輦輩輪輬輭輮輯輲輳輴輵輶輷輸輹輻輾輿轀轂轃轄轅轆轇轈轉轍轎轏轐轑轒轓轔轕轖轗轘轙轚轛轝轞轟轠轡轢轤车轧轨轩轪轫转轭轮软轰轱轲轳轴轵轶轷轸轹轺轻轼载轾轿辀辁辂较辄辅辆辇辈辉辊辋辌辍辎辏辐辑辒输辔辕辖辗辘辙辚辛辜辞辟辠辢辣辥辦辧辨辩辫辭辮辯辰辱農辴辵边辺辻込辽达辿迁迂迄迅迆过迈迉迋迍迎运近迒迓返迕迖还这进远违连迟迡迢迣迤迥迦迨迩迪迫迭迮述迳迴迵迷迸迹迺迻追迾迿退送适逃逄逅逆逈选逊逋逌逍逎透逐逑递逓途逕逖逗這通逛逜逝逞速造逡逢連逤逦逨逩逭逮逯週進逴逵逶逸逹逺逻逼逾逿遁遂遄遅遆遇遊運遍過遏遐遑遒道達違遗遘遙遛遜遝遞遠遡遢遣遥遨適遭遮遯遰遲遳遴遵遶遷選遹遺遻遼遽遾避邀邁邂邃還邅邆邇邈邉邊邋邍邏邐邑邓邔邕邗邘邙邛邝邞邟邠邡邢那邥邦邧邨邪邬邮邯邰邱邲邳邴邵邶邸邹邺邻邽邾邿郁郃郄郅郇郈郊郋郎郏郐郑郓郔郕郖郗郘郙郚郛郜郝郞郟郠郡郢郣郤郥郦郧部郪郫郭郯郰郱郲郳郴郵郷郸郹郺郻郼都郾郿鄀鄁鄂鄃鄄鄅鄆鄇鄈鄉鄋鄌鄍鄎鄏鄐鄑鄒鄔鄖鄗鄘鄙鄚鄛鄜鄝鄞鄟鄠鄡鄢鄣鄤鄦鄧鄨鄩鄪鄫鄬鄭鄮鄯鄰鄱鄲鄳鄴鄵鄶鄸鄹鄺鄻鄾酀酁酂酃酄酅酆酇酈酉酊酋酌配酎酏酐酒酓酔酖酗酘酚酝酞酟酡酢酣酤酥酦酧酨酩酪酬酮酯酰酱酲酳酴酵酶酷酸酹酺酼酽酾酿醁醂醃醅醆醇醉醊醋醌醍醐醑醒醓醕醖醗醙醚醛醜醞醟醢醣醤醥醧醨醪醫醬醭醮醯醰醱醲醳醴醵醷醸醹醺醻醼醽醾醿釀釁釂釃釅釆采釈釉释釋里重野量釐金釓釔釕釗釘釙釚釜針釢釣釤釦釧釩釪釬釭釱釳釴釵釷釸釹釺釽釿鈀鈁鈃鈄鈆鈇鈈鈉鈊鈌鈍鈎鈏鈐鈑鈒鈔鈕鈖鈚鈞鈢鈣鈥鈦鈧鈫鈮鈰鈳鈴鈶鈷鈸鈹鈺鈽鈾鈿鉀鉄鉅鉆鉇鉈鉉鉊鉋鉌鉍鉏鉑鉒鉓鉕鉗鉚鉛鉞鉟鉠鉢鉤鉥鉦鉧鉬鉭鉮鉯鉱鉲鉴鉶鉷鉸鉹鉺鉻鉼鉾鉿銀銂銃銅銈銋銍銎銑銓銔銕銖銗銘銚銛銜銠銣銤銥銦銨銩銪銫銬銭銮銱銲銳銶銷銹銻銼銾鋀鋁鋂鋃鋄鋅鋆鋇鋈鋊鋋鋌鋏鋐鋒鋕鋗鋘鋙鋝鋟鋡鋤鋥鋦鋨鋩鋪鋭鋮鋯鋰鋱鋲鋳鋶鋸鋹鋺鋼錀錁錄錆錈錍錎錏錐錒錕錘錙錚錛錞錟錠錡錢錣錦錧錨錫錬錭錮錯録錳錵錶錸錻錼錾鍀鍁鍆鍇鍈鍉鍊鍋鍌鍍鍏鍐鍑鍔鍖鍘鍚鍛鍜鍝鍟鍠鍤鍥鍩鍪鍬鍭鍮鍰鍱鍳鍵鍶鍷鍺鍼鍾鎂鎃鎅鎉鎊鎌鎎鎏鎒鎓鎔鎖鎗鎘鎚鎛鎝鎞鎡鎢鎣鎦鎧鎩鎪鎬鎭鎮鎯鎰鎳鎵鎷鎸鎹鎻鏂鏃鏇鏈鏊鏌鏍鏏鏐鏑鏓鏖鏗鏘鏙鏚鏜鏝鏞鏟鏡鏢鏤鏦鏨鏬鏰鏳鏴鏵鏷鏹鏺鏻鏽鏾鐀鐃鐇鐈鐋鐍鐎鐏鐐鐓鐔鐕鐖鐘鐙鐚鐝鐠鐡鐤鐦鐧鐨鐫鐬鐭鐮鐯鐲鐳鐵鐶鐸鐺鐻鐽鐾鐿鑀鑁鑂鑄鑅鑊鑋鑌鑐鑑鑒鑔鑕鑚鑛鑞鑠鑢鑣鑤鑨鑪鑫鑭鑮鑯鑰鑱鑲鑴鑵鑷鑹鑼鑽鑾鑿钀钁钂钃钆钇针钉钊钋钌钍钎钏钐钒钓钔钕钖钗钘钙钚钛钜钝钞钟钠钡钢钣钤钥钦钧钨钩钪钫钬钭钮钯钰钱钲钳钴钵钷钹钺钻钼钽钾钿铀铁铂铃铄铅铆铈铉铊铋铌铍铎铏铐铑铒铕铖铗铘铙铚铛铜铝铞铟铠铡铢铣铤铥铧铨铩铪铫铬铭铮铯铰铱铲铳铴铵银铷铸铹铺铻铼铽链铿销锁锂锃锄锅锆锇锈锉锊锋锌锍锎锏锐锑锒锓锔锕锖锗锘错锚锛锜锝锞锟锡锢锣锤锥锦锧锨锩锪锫锬锭键锯锰锱锲锳锴锵锶锷锸锹锺锻锼锽锾锿镀镁镂镃镄镅镆镇镈镉镊镋镌镍镎镏镐镑镒镓镔镕镖镗镘镚镛镜镝镞镠镡镢镣镤镥镦镧镨镩镪镫镬镭镮镯镰镱镲镳镴镵镶長镺镻镼镽长門閂閃閆閇閈閉開閌閍閎閏閑閒間閔閘閜閞閟閡関閣閤閥閦閧閨閩閫閬閭閰閱閲閵閶閷閹閺閻閼閽閾閿闀闃闅闆闇闈闉闊闋闌闍闐闑闒闓闔闕闖闘闚闛關闞闟闠闡闢闤闥门闩闪闫闭问闯闰闱闲闳间闵闶闷闸闹闺闻闼闽闾闿阀阁阂阃阄阅阆阇阈阉阊阋阌阍阎阏阐阑阒阔阕阖阗阘阙阚阜阝阞队阠阡阢阤阨阪阬阭阮阯阰阱防阳阴阵阶阸阹阺阻阼阽阿陀陁陂附际陆陇陈陉陊陋陌降陎陏限陑陓陔陕陗陘陛陜陝陞陟陡院陣除陥陧陨险陪陫陬陭陰陲陳陴陵陶陷陸険陻陼陽陾隃隄隅隆隇隈隉隊隋隍階随隐隑隒隓隔隕隗隘隙際障隞隠隣隤隧隨隩險隮隰隱隲隳隴隶隷隸隹隺隻隼隽难隿雀雁雂雄雅集雇雈雉雊雋雌雍雎雏雑雒雓雔雕雖雗雘雙雚雛雜雝雞雟雠離難雨雩雪雫雯雰雱雲雳零雷雹雺電雽雾雿需霁霂霄霅霆震霈霉霊霋霍霎霏霐霑霒霓霖霙霜霞霠霢霣霤霧霨霩霪霫霬霭霮霰露霵霶霸霹霽霾霿靁靂靃靄靆靈靉靑青靓靖静靚靛靜非靠靡面靥靦靨革靪靫靬靭靮靰靱靲靳靴靶靷靸靺靻靼靽靾靿鞀鞁鞂鞃鞄鞅鞆鞈鞊鞋鞌鞍鞎鞏鞑鞒鞔鞗鞘鞙鞚鞜鞞鞠鞡鞣鞤鞥鞦鞧鞨鞪鞫鞬鞭鞮鞯鞲鞳鞴鞵鞶鞷鞹鞻鞽鞾鞿韁韂韃韄韅韆韇韉韋韌韍韎韏韐韑韓韔韕韗韙韜韝韞韟韠韡韣韥韦韧韨韩韪韫韬韭韮韰韱音韵韶韹韺韻韽韾響頁頂頃頄項順頇須頊頌頍頎頏預頑頒頓頔頖頗領頚頜頝頞頠頡頤頦頨頩頫頬頭頯頰頲頴頵頷頸頹頻頼頽顁顄顅顆顇顈顉顊顋題額顎顏顐顑顒顓顔顕顗願顙顛顜顝類顟顠顢顣顤顥顦顧顩顪顫顯顰顱顲顳顴页顶顷顸项顺须顼顽顾顿颀颁颂颃预颅领颇颈颉颊颋颌颍颎颏颐频颓颔颖颗题颙颚颛颜额颞颟颠颡颢颤颥颦颧風颬颭颮颯颱颲颳颶颸颺颻颼颽颾颿飁飂飄飆飈飉飋飌风飏飐飑飒飓飔飕飗飘飙飛飜飞食飡飢飣飥飧飨飩飪飫飬飭飮飯飱飲飴飶飺飼飽飾飿餀餂餃餄餅餈餉養餌餍餎餏餐餑餒餓餔餕餖餗餘餚餛餜餝餞餟餠餡餤餥餧館餪餫餬餭餮餯餰餱餲餳餵餷餹餺餻餼餽餾餿饁饂饃饅饇饈饉饊饋饌饍饎饐饑饒饔饕饗饘饙饛饜饞饟饡饢饥饧饨饩饪饫饬饭饮饯饰饱饲饳饴饵饶饷饸饹饺饻饼饽饿馁馃馄馅馆馇馈馉馊馋馌馍馏馐馑馒馓馔馕首馗馘香馛馝馞馣馥馦馧馨馬馭馮馯馰馱馲馳馴馵馹馺馻馼馽駁駂駃駄駅駆駈駉駍駎駏駐駑駒駓駔駕駖駘駙駛駜駝駟駡駢駣駤駥駩駪駬駭駮駰駱駴駶駷駸駹駺駻駼駽駾駿騁騂騃騄騅騇騉騊騋騌騍騎騏騑騒験騕騖騙騚騛騜騝騞騠騢騣騤騥騧騨騩騪騫騬騭騮騰騱騲騴騵騶騷騸騹騺騽騾驀驁驂驃驄驅驆驈驉驊驌驍驎驏驒驓驔驕驖驗驚驛驞驟驠驢驤驥驦驨驩驪驫马驭驮驯驰驱驲驳驴驵驶驷驸驹驺驻驼驽驾驿骀骁骂骃骄骅骆骇骈骉骊骋验骍骎骏骐骑骒骓骕骖骗骘骙骚骛骜骝骞骟骠骡骢骣骤骥骦骧骨骫骭骯骰骱骴骶骷骸骹骺骼骽骾骿髀髁髂髃髄髅髆髇髈髊髋髌髍髎髏髐髑髒髓體髕髖高髙髟髡髢髣髥髦髧髪髫髬髭髮髯髲髳髴髶髷髹髺髻髼髽髾鬀鬁鬃鬄鬅鬆鬈鬉鬊鬋鬌鬍鬎鬏鬐鬒鬓鬕鬖鬗鬘鬙鬚鬞鬟鬠鬢鬣鬤鬥鬧鬨鬩鬪鬫鬬鬮鬯鬱鬲鬳鬵鬶鬷鬹鬺鬻鬼鬾鬿魁魂魃魄魅魆魇魈魉魊魋魌魍魎魏魑魔魕魖魘魚魛魞魟魠魡魣魦魧魨魩魮魯魱魴魵魶魷魻魼魾鮀鮁鮂鮃鮄鮅鮆鮇鮊鮋鮍鮎鮐鮑鮒鮓鮗鮚鮛鮜鮝鮞鮟鮠鮡鮣鮤鮥鮦鮨鮪鮫鮭鮮鮰鮵鮸鮹鮻鮿鯀鯁鯃鯆鯇鯈鯉鯊鯏鯒鯓鯔鯕鯖鯗鯙鯛鯝鯞鯠鯡鯢鯤鯦鯧鯨鯪鯫鯬鯭鯮鯰鯱鯶鯷鯸鯻鯽鯿鰂鰃鰅鰆鰈鰉鰋鰌鰍鰏鰐鰒鰓鰕鰗鰛鰜鰝鰟鰣鰤鰥鰧鰨鰩鰫鰬鰭鰮鰯鰰鰱鰲鰳鰴鰶鰷鰹鰺鰻鰼鰾鰿鱀鱁鱂鱄鱅鱆鱇鱈鱉鱊鱋鱍鱐鱒鱓鱔鱕鱖鱗鱘鱚鱝鱞鱟鱠鱢鱣鱥鱦鱧鱨鱬鱭鱮鱰鱲鱳鱴鱵鱷鱸鱹鱺鱻鱼鱽鱾鱿鲀鲁鲂鲃鲅鲆鲇鲈鲉鲊鲋鲌鲍鲎鲏鲐鲑鲒鲔鲕鲖鲗鲘鲙鲚鲛鲜鲝鲞鲟鲠鲡鲢鲣鲤鲥鲦鲧鲨鲩鲪鲫鲬鲭鲮鲯鲰鲱鲲鲳鲴鲵鲷鲸鲹鲺鲻鲼鲽鲾鲿鳀鳁鳂鳃鳄鳅鳇鳈鳉鳊鳌鳍鳎鳏鳐鳑鳒鳓鳔鳕鳖鳗鳘鳙鳚鳛鳜鳝鳞鳟鳠鳡鳢鳣鳤鳥鳦鳧鳩鳪鳬鳭鳯鳰鳱鳲鳳鳴鳵鳶鳷鳹鳺鳻鳼鳽鳾鳿鴀鴃鴅鴆鴇鴈鴉鴎鴐鴒鴓鴔鴕鴗鴘鴛鴝鴞鴟鴠鴡鴢鴣鴥鴦鴨鴩鴫鴮鴯鴰鴱鴳鴴鴶鴷鴸鴻鴽鴾鴿鵀鵁鵂鵃鵄鵅鵊鵋鵌鵎鵏鵐鵑鵒鵓鵔鵖鵗鵙鵛鵜鵝鵞鵟鵠鵡鵧鵩鵪鵫鵬鵮鵯鵰鵱鵲鵳鵴鵵鵷鵸鵹鵺鵻鵼鵽鵾鶀鶁鶂鶄鶅鶆鶇鶈鶉鶊鶋鶌鶏鶐鶒鶓鶔鶖鶗鶘鶙鶚鶛鶜鶝鶞鶟鶠鶡鶢鶣鶤鶥鶨鶩鶪鶬鶭鶯鶱鶲鶴鶵鶶鶷鶹鶺鶻鶼鶾鶿鷀鷁鷂鷃鷄鷅鷇鷈鷉鷊鷋鷎鷏鷐鷑鷒鷓鷕鷖鷗鷘鷙鷚鷛鷜鷞鷟鷡鷢鷣鷤鷥鷦鷩鷫鷮鷯鷰鷲鷳鷵鷶鷷鷸鷹鷺鷻鷽鷾鷿鸀鸁鸂鸃鸄鸅鸆鸇鸉鸊鸋鸌鸍鸏鸐鸑鸒鸓鸔鸕鸗鸘鸙鸚鸛鸜鸝鸞鸟鸠鸡鸢鸣鸤鸥鸦鸧鸨鸩鸪鸫鸬鸭鸮鸯鸰鸱鸲鸳鸵鸶鸷鸸鸹鸺鸻鸼鸽鸾鸿鹀鹁鹂鹃鹄鹅鹆鹇鹈鹉鹊鹋鹌鹍鹎鹏鹐鹑鹒鹔鹕鹖鹗鹘鹙鹚鹛鹜鹝鹞鹟鹠鹡鹢鹣鹤鹦鹧鹨鹩鹪鹫鹬鹭鹮鹯鹰鹱鹲鹳鹴鹵鹸鹹鹺鹻鹼鹽鹾鹿麀麁麂麃麅麆麇麈麉麊麋麌麎麐麑麒麓麔麖麗麙麚麛麜麝麞麟麠麡麤麥麦麧麩麪麮麯麰麴麵麶麷麸麹麺麻麼麽麾麿黀黁黂黃黄黇黈黉黌黍黎黏黐黑黒黓黔黕黖默黙黚黛黜黝點黟黠黡黢黤黥黧黨黩黪黫黭黮黯黰黲黳黴黵黶黷黹黺黻黼黽黾黿鼀鼁鼂鼃鼆鼇鼈鼉鼊鼋鼍鼎鼏鼐鼒鼓鼕鼖鼗鼘鼙鼛鼜鼠鼢鼣鼤鼥鼨鼩鼪鼫鼬鼭鼮鼯鼰鼱鼲鼳鼴鼵鼶鼷鼸鼹鼻鼽鼾齀齁齂齃齇齉齊齋齌齍齎齏齐齑齒齔齕齖齗齘齙齛齜齝齞齟齠齡齢齣齤齥齦齧齩齪齫齬齮齯齰齱齲齴齵齶齷齸齹齺齻齾齿龀龁龂龃龄龅龆龇龈龉龊龋龌龍龎龐龑龒龔龕龗龙龚龛龜龝龟龠龢龤鿍鿎鿏郎凉︽︾﹏﹥﹪！＃％＆（）＋，－．／０１２３４５６７８９：；＝？＠ＡＢＣＤＥＦＧＨＩＪＫＬＭＮＯＰＱＲＳＴＵＶＷＸＹＺａｂｃｄｅｆｇｈｉｊｋｌｍｎｏｐｑｒｓｔｕｖｗｘｙｚ｜～･ｦｧｨｩｪｫｬｭｮｯｱｲｳｴｵｶｷｸｹｺｻｼｽｾｿﾀﾁﾂﾃﾄﾅﾆﾇﾈﾉﾊﾋﾌﾍﾎﾏﾐﾑﾒﾓﾔﾕﾖﾗﾘﾙﾚﾛﾜﾝﾞﾟ￥𠅤𠙶𠳐𠴰𡎚𡐓𡑍𡚁𣇉𣗋𣲗𣲘𣸣𤋮𤧛𤩽𤫉𤺥𥔲𥕢𥖨𥻗𦈡𦍑𦒍𦙶𦝼𦭜𦰡𧿹𨐈𨙸𨚕𨟠𨭉𨱇𨱏𨱑𨱔𨺙𩓐𩽾𩾃𩾌𪟝𪣻𪤗𪨰𪨶𪩘𪾢𫄧𫄨𫄷𫄸𫇭𫌀𫍣𫍯𫍲𫍽𫐄𫐐𫐓𫑡𫓧𫓯𫓶𫓹𫔍𫔎𫔶𫖮𫖯𫖳𫗧𫗴𫘜𫘝𫘦𫘧𫘨𫘪𫘬𫚕𫚖𫚭𫛭𫞩𫟅𫟦𫟹𫟼𫠆𫠊𫠜𫢸𫫇𫭟𫭢𫭼𫮃𫰛𫵷𫶇𫷷𫸩𬀩𬀪𬂩𬃊𬇕𬇙𬇹𬉼𬊈𬊤𬌗𬍛𬍡𬍤𬒈𬒔𬒗𬕂𬘓𬘘𬘡𬘩𬘫𬘬𬘭𬘯𬙂𬙊𬙋𬜬𬜯𬞟𬟁𬟽𬣙𬣞𬣡𬣳𬤇𬤊𬤝𬨂𬨎𬩽𬪩𬬩𬬭𬬮𬬱𬬸𬬹𬬻𬬿𬭁𬭊𬭎𬭚𬭛𬭤𬭩𬭬𬭯𬭳𬭶𬭸𬭼𬮱𬮿𬯀𬯎𬱖𬱟𬳵𬳶𬳽𬳿𬴂𬴃𬴊𬶋𬶍𬶏𬶐𬶟𬶠𬶨𬶭𬶮𬷕𬸘𬸚𬸣𬸦𬸪𬹼𬺈𬺓🀄🃏🅰🅱🅾🅿🆎🆑🆒🆓🆔🆕🆖🆗🆘🆙🆚🇨🇩🇪🇫🇬🇮🇯🇰🇷🇺🈁🌀🌁🌂🌃🌄🌅🌆🌇🌈🌉🌊🌋🌌🌍🌎🌏🌐🌑🌒🌓🌔🌕🌖🌗🌘🌙🌚🌛🌜🌝🌞🌟🌠🌰🌱🌲🌳🌴🌵🌷🌸🌹🌺🌻🌼🌽🌾🌿🍀🍁🍂🍃🍄🍅🍆🍇🍈🍉🍊🍋🍌🍍🍎🍏🍐🍑🍒🍓🍔🍕🍖🍗🍘🍙🍚🍛🍜🍝🍞🍟🍠🍡🍢🍣🍤🍥🍦🍧🍨🍩🍪🍫🍬🍭🍮🍯🍰🍱🍲🍳🍴🍵🍶🍷🍸🍹🍺🍻🍼🎀🎁🎂🎃🎄🎅🎆🎇🎈🎉🎊🎋🎌🎍🎎🎏🎐🎑🎒🎓🎠🎡🎢🎣🎤🎥🎦🎧🎨🎩🎪🎫🎬🎭🎮🎯🎰🎱🎲🎳🎴🎵🎶🎷🎸🎹🎺🎻🎼🎽🎾🎿🏀🏁🏂🏃🏄🏆🏇🏈🏉🏊🏠🏡🏢🏣🏤🏥🏦🏧🏨🏩🏪🏫🏬🏭🏮🏯🏰🐀🐁🐂🐃🐄🐅🐆🐇🐈🐉🐊🐋🐌🐍🐎🐏🐐🐑🐒🐓🐔🐕🐖🐗🐘🐙🐚🐛🐜🐝🐞🐟🐠🐡🐢🐣🐤🐥🐦🐧🐨🐩🐪🐫🐬🐭🐮🐯🐰🐱🐲🐳🐴🐵🐶🐷🐸🐹🐺🐻🐼🐽🐾👀👂👃👄👅👆👇👈👉👊👋👌👍👎👏👐👑👒👓👔👕👖👗👘👙👚👛👜👝👞👟👠👡👢👣👤👥👦👧👨👩👪👫👬👭👮👯👰👱👲👳👴👵👶👷👸👹👺👻👼👽👾👿💀💁💂💃💄💅💆💇💈💉💊💋💌💍💎💏💐💑💒💓💔💕💖💗💘💙💚💛💜💝💞💟💠💡💢💣💤💥💦💧💨💩💪💫💬💭💮💯💰💲💳💴💵💶💷💸💹💺💻💼💽💾💿📀📁📂📃📄📅📆📇📈📉📊📋📌📍📎📏📐📑📒📓📔📕📖📗📘📙📚📛📜📝📞📟📠📡📢📣📤📥📦📧📨📩📪📫📬📭📮📯📰📱📲📳📴📵📶📷📹📺📻📼🔀🔁🔂🔃🔄🔅🔆🔇🔉🔊🔋🔌🔍🔎🔏🔐🔑🔒🔓🔔🔕🔖🔗🔘🔙🔚🔛🔜🔝🔞🔟🔠🔡🔢🔣🔤🔥🔦🔧🔨🔩🔪🔫🔬🔭🔮🔯🔰🔱🔲🔳🔴🔵🔶🔷🔸🔹🔺🔻🔼🔽🕐🕑🕒🕓🕔🕕🕖🕗🕘🕙🕚🕛🕜🕝🕞🕟🕠🕡🕢🕣🕤🕥🕦🕧🗻🗼🗽🗾🗿😀😁😂😃😄😅😆😇😈😉😊😋😌😍😎😏😐😑😒😓😔😕😖😗😘😙😚😛😜😝😞😟😠😡😢😣😤😥😦😧😨😩😪😫😬😭😮😯😰😱😲😳😴😵😶😷😸😹😺😻😼😽😾😿🙀🙅🙆🙇🙈🙉🙊🙋🙌🙍🙎🙏🚀🚁🚂🚃🚄🚅🚆🚇🚈🚉🚊🚌🚍🚎🚏🚐🚑🚒🚓🚔🚕🚖🚗🚘🚙🚚🚛🚜🚝🚞🚟🚠🚡🚢🚣🚤🚥🚦🚧🚨🚩🚪🚫🚬🚭🚮🚯🚰🚱🚲🚳🚴🚵🚶🚷🚸🚹🚺🚻🚼🚽🚾🚿🛀🛁🛂🛃🛄🛅 ';
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
new file mode 100644
index 0000000000..0c2dc3d283
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
@@ -0,0 +1,152 @@
+// Built-in detector box-extraction strategies for the OCR pipeline. Each is a
+// TextBoxExtractor for one detector architecture: it calls that architecture's
+// native decoder and reshapes the flat output into quads. The pipeline stays
+// model-agnostic — it just invokes OcrOptions.extractBoxes; the presets in
+// models.ts wire the zero-config built-ins, a model with non-standard
+// thresholds uses the make* factories, and a new architecture plugs in by
+// supplying its own conforming function.
+
+import { rnexecutorchJsi } from '../../../../native/bridge';
+import type { Tensor } from '../../../../core/tensor';
+import { quadsFromFlat, type Quad } from '../../ops/quad';
+
+/**
+ * A detector's box-extraction strategy: turns a model's raw `detect_<S>` output
+ * tensors into oriented {@link Quad}s in detector-input pixel space. A model plugs a
+ * new detector into the OCR pipeline by supplying a function of this type (the
+ * built-ins below, or its own). MUST be a worklet.
+ * @category Types
+ * @param outputs The model's `detect_<S>` output tensors, in order (owned by the
+ * pipeline — do not dispose).
+ * @param side The snapped square detector side `S` (input is `S × S` letterboxed).
+ * @param charLevel Emit one box per glyph instead of grouped lines; strategies
+ * without a char-level mode ignore it.
+ * @returns Oriented quads (TL, TR, BR, BL) in detector-input pixel space.
+ */
+export type TextBoxExtractor = (
+  outputs: readonly Tensor[],
+  side: number,
+  charLevel: boolean
+) => Quad[];
+
+/**
+ * Threshold overrides for {@link makeCraftExtractBoxes}. Any omitted field keeps
+ * the CRAFT default.
+ * @category Types
+ */
+export type CraftExtractorOptions = {
+  /** Region-heatmap binarization threshold. Defaults to 0.4. */
+  readonly textThreshold?: number;
+  /** Affinity-heatmap (glyph linking) threshold. Defaults to 0.4. */
+  readonly linkThreshold?: number;
+  /** Minimum component peak score to keep a box. Defaults to 0.7. */
+  readonly lowTextThreshold?: number;
+};
+
+/**
+ * Threshold overrides for {@link makeDbnetExtractBoxes}. Any omitted field keeps
+ * the DBNet default.
+ * @category Types
+ */
+export type DbnetExtractorOptions = {
+  /** Probability-map binarization threshold. Defaults to 0.3. */
+  readonly binThreshold?: number;
+  /** Minimum mean in-contour probability to keep a box. Defaults to 0.6. */
+  readonly boxThreshold?: number;
+  /** Box expansion (unclip) ratio. Defaults to 1.5. */
+  readonly unclipRatio?: number;
+  /** Minimum box side in pixels. Defaults to 3. */
+  readonly minBoxSide?: number;
+  /** Maximum contour candidates considered. Defaults to 1000. */
+  readonly maxCandidates?: number;
+};
+
+// CRAFT region+affinity heatmap thresholds — stable across models, the defaults.
+const CRAFT_TEXT_THRESHOLD = 0.4;
+const CRAFT_LINK_THRESHOLD = 0.4;
+const CRAFT_LOW_TEXT_THRESHOLD = 0.7;
+
+// DBNet probability-map thresholds — stable across models, the defaults.
+const DBNET_BIN_THRESHOLD = 0.3;
+const DBNET_BOX_THRESHOLD = 0.6;
+const DBNET_UNCLIP_RATIO = 1.5;
+const DBNET_MIN_BOX_SIDE = 3;
+const DBNET_MAX_CANDIDATES = 1000;
+
+/**
+ * Builds a CRAFT {@link TextBoxExtractor} with custom thresholds: groups the
+ * half-resolution region+affinity heatmap (`outputs[0]` is the `[1,Hd,Wd,2]`
+ * heatmap) into oriented text-line quads, or per-glyph boxes when `charLevel`.
+ * For the standard thresholds use the ready-made {@link craftExtractBoxes}.
+ * @category Typescript API
+ * @param overrides Threshold overrides; omitted fields keep the CRAFT defaults.
+ * @returns A {@link TextBoxExtractor} to assign to `OcrOptions.extractBoxes`.
+ */
+export function makeCraftExtractBoxes(overrides?: CraftExtractorOptions): TextBoxExtractor {
+  const textThreshold = overrides?.textThreshold ?? CRAFT_TEXT_THRESHOLD;
+  const linkThreshold = overrides?.linkThreshold ?? CRAFT_LINK_THRESHOLD;
+  const lowTextThreshold = overrides?.lowTextThreshold ?? CRAFT_LOW_TEXT_THRESHOLD;
+  return (outputs, side, charLevel) => {
+    'worklet';
+    // The half-resolution heatmap requires an even detector side; the pipeline is
+    // architecture-agnostic and can't check this, so the strategy does.
+    if (side % 2 !== 0) {
+      throw new Error(
+        'OCR: every CRAFT detect bucket side must be even (half-resolution heatmap).'
+      );
+    }
+    const flat = rnexecutorchJsi.cv.extractCraftTextBoxes(outputs[0]!, {
+      textThreshold,
+      linkThreshold,
+      lowTextThreshold,
+      targetHeight: side,
+      charLevel,
+    }) as number[];
+    return quadsFromFlat(flat);
+  };
+}
+
+/**
+ * Builds a DBNet {@link TextBoxExtractor} with custom thresholds: thresholds and
+ * unclips the probability map (`outputs[0]` is the `[1,1,H,W]` post-sigmoid prob
+ * map) into oriented text quads. It decodes at full resolution with no char-level
+ * mode, so the extractor uses neither `side` nor `charLevel`. For the standard
+ * thresholds use the ready-made {@link dbnetExtractBoxes}.
+ * @category Typescript API
+ * @param overrides Threshold overrides; omitted fields keep the DBNet defaults.
+ * @returns A {@link TextBoxExtractor} to assign to `OcrOptions.extractBoxes`.
+ */
+export function makeDbnetExtractBoxes(overrides?: DbnetExtractorOptions): TextBoxExtractor {
+  const binThreshold = overrides?.binThreshold ?? DBNET_BIN_THRESHOLD;
+  const boxThreshold = overrides?.boxThreshold ?? DBNET_BOX_THRESHOLD;
+  const unclipRatio = overrides?.unclipRatio ?? DBNET_UNCLIP_RATIO;
+  const minBoxSide = overrides?.minBoxSide ?? DBNET_MIN_BOX_SIDE;
+  const maxCandidates = overrides?.maxCandidates ?? DBNET_MAX_CANDIDATES;
+  return (outputs) => {
+    'worklet';
+    const flat = rnexecutorchJsi.cv.extractDbnetTextBoxes(outputs[0]!, {
+      binThreshold,
+      boxThreshold,
+      unclipRatio,
+      minBoxSide,
+      maxCandidates,
+    }) as number[];
+    return quadsFromFlat(flat);
+  };
+}
+
+/**
+ * Built-in CRAFT {@link TextBoxExtractor} with the standard thresholds. Assign to
+ * `OcrOptions.extractBoxes` for a CRAFT-family model (the detector side must be
+ * even), or build a custom-threshold variant with {@link makeCraftExtractBoxes}.
+ * @category Typescript API
+ */
+export const craftExtractBoxes: TextBoxExtractor = makeCraftExtractBoxes();
+
+/**
+ * Built-in DBNet {@link TextBoxExtractor} with the standard thresholds. Assign to
+ * `OcrOptions.extractBoxes` for a DBNet-family model, or build a
+ * custom-threshold variant with {@link makeDbnetExtractBoxes}.
+ * @category Typescript API
+ */
+export const dbnetExtractBoxes: TextBoxExtractor = makeDbnetExtractBoxes();
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
index 8d7f8cbbcd..39fe40c657 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
@@ -1,141 +1,125 @@
 import type { ImageBuffer } from '../../image';
 import { FORMAT_CHANNELS } from '../../ops/image';
 import { boundingBoxOf, type BoundingBox } from '../../ops/boxes';
-import type { OCRDetection } from '../ocr';
+import type { OcrDetection } from '../ocr';
 
-// Crops an axis-aligned region out of an ImageBuffer (pure pixel slice, same
-// format). Used to feed a layout region (e.g. a table) to another model.
+/**
+ * Crops an axis-aligned region out of an image as a plain pixel slice (same format
+ * and layout). Used to feed a layout region to another model.
+ * @category Typescript API
+ * @param input The source image.
+ * @param bbox The crop region, in `xyxy` pixels.
+ * @returns The cropped image.
+ */
 export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>): ImageBuffer {
   'worklet';
   const { data, width, height, format } = input;
-  const ch = FORMAT_CHANNELS[format];
+  const channels = FORMAT_CHANNELS[format];
   const x0 = Math.max(0, Math.min(Math.round(bbox.xmin), width));
   const y0 = Math.max(0, Math.min(Math.round(bbox.ymin), height));
   const x1 = Math.max(0, Math.min(Math.round(bbox.xmax), width));
   const y1 = Math.max(0, Math.min(Math.round(bbox.ymax), height));
-  const cw = Math.max(1, x1 - x0);
-  const chh = Math.max(1, y1 - y0);
-  const out = new Uint8Array(cw * chh * ch);
-  for (let y = 0; y < chh; y++) {
-    const srcStart = ((y0 + y) * width + x0) * ch;
-    out.set(data.subarray(srcStart, srcStart + cw * ch), y * cw * ch);
-  }
-  return { data: out, width: cw, height: chh, format, layout: input.layout };
-}
-
-/**
- * Index of the maximum value in `arr[offset, offset+len)` (single pass, no
- * allocation). Shared by the layout / orientation / table decoders.
- * @category Typescript API
- * @param arr The array to scan.
- * @param offset The start index of the window.
- * @param len The window length.
- * @returns The index (relative to `offset`) of the maximum value.
- */
-export function argmaxRange(arr: ArrayLike<number>, offset: number, len: number): number {
-  'worklet';
-  let idx = 0;
-  let best = arr[offset]!;
-  for (let i = 1; i < len; i++) {
-    const v = arr[offset + i]!;
-    if (v > best) {
-      best = v;
-      idx = i;
-    }
+  const cropWidth = Math.max(1, x1 - x0);
+  const cropHeight = Math.max(1, y1 - y0);
+  const out = new Uint8Array(cropWidth * cropHeight * channels);
+  for (let y = 0; y < cropHeight; y++) {
+    const rowStart = ((y0 + y) * width + x0) * channels;
+    out.set(data.subarray(rowStart, rowStart + cropWidth * channels), y * cropWidth * channels);
   }
-  return idx;
+  return { data: out, width: cropWidth, height: cropHeight, format, layout: input.layout };
 }
 
-// Assigns each value to one of `k` ordered groups by cutting the sorted values at
-// the (k-1) largest gaps, then returns the MEAN of each group (its center). Never
-// a fixed gap, which would merge dense rows.
-function splitIntoK(values: readonly number[], k: number): number[] {
+// Groups values into `k` ordered clusters by cutting the sorted values at their
+// (k-1) widest gaps, then returns each cluster's mean. Cutting at the widest gaps
+// (rather than at fixed intervals) keeps dense rows/columns together.
+function clusterCentersByGaps(values: readonly number[], k: number): number[] {
   'worklet';
-  const v = [...values].sort((a, b) => a - b);
-  if (v.length <= k) {
-    return v;
+  const sorted = [...values].sort((a, b) => a - b);
+  if (sorted.length <= k) {
+    return sorted;
   }
-  const cuts = v
-    .map((x, i) => ({ i, g: i ? x - v[i - 1]! : -1 }))
-    .slice(1)
-    .sort((a, b) => b.g - a.g)
+  // Rank the interior gaps (gap i sits between sorted[i-1] and sorted[i]) and take
+  // the k-1 widest as cut points, restored to ascending order.
+  const gaps = sorted.slice(1).map((value, i) => ({ at: i + 1, size: value - sorted[i]! }));
+  gaps.sort((a, b) => b.size - a.size);
+  const cuts = gaps
     .slice(0, k - 1)
-    .map((o) => o.i)
+    .map((gap) => gap.at)
     .sort((a, b) => a - b);
+  // Average each [prev, cut) span into its center.
   const centers: number[] = [];
   let prev = 0;
-  for (const cut of [...cuts, v.length]) {
-    const grp = v.slice(prev, cut);
-    centers.push(grp.reduce((s, x) => s + x, 0) / grp.length);
+  for (const cut of [...cuts, sorted.length]) {
+    const group = sorted.slice(prev, cut);
+    centers.push(group.reduce((sum, value) => sum + value, 0) / group.length);
     prev = cut;
   }
   return centers;
 }
 
-// Index of the center in `cs` closest to `x`.
-function nearest(x: number, cs: readonly number[]): number {
+// Index of the center in `centers` nearest to `value`.
+function nearestIndex(value: number, centers: readonly number[]): number {
   'worklet';
-  let b = 0;
-  for (let j = 1; j < cs.length; j++) {
-    if (Math.abs(x - cs[j]!) < Math.abs(x - cs[b]!)) {
-      b = j;
+  let best = 0;
+  for (let i = 1; i < centers.length; i++) {
+    if (Math.abs(value - centers[i]!) < Math.abs(value - centers[best]!)) {
+      best = i;
     }
   }
-  return b;
+  return best;
 }
 
 /**
- * Reconstructs a table as filled HTML from the SLANet structure skeleton and the
- * region's OCR lines. Grid size comes from the structure (`<tr>` count, max `<td>`
- * per row); box centers are split into that many rows/cols at the largest gaps and
- * each box placed in its nearest cell (global column centers keep columns aligned).
- * Falls back to a document-order fill when there is no grid.
+ * Fills a table-structure HTML skeleton with a region's OCR lines. The grid size
+ * comes from the skeleton (row count, and the widest row's cell count); each
+ * line's box center is assigned to its nearest row and column cluster, so shared
+ * column centers keep columns aligned. Falls back to a document-order fill when the
+ * skeleton has no grid.
  *
- * Geometric alignment only — dense bottom rows can misplace a value, since the
- * structure export has no cell boxes.
+ * Alignment is geometric only — dense rows can misplace a value, since the
+ * skeleton carries no per-cell coordinates.
  * @category Typescript API
- * @param html The SLANet structure HTML skeleton (gives the grid size).
- * @param lines The table region's OCR lines (with page-space quads).
+ * @param html The structure HTML skeleton (empty cells).
+ * @param lines The region's OCR lines, with page-space quads.
  * @returns A `<table>` with each cell filled by its nearest-assigned text.
  */
-export function fillTableCells(html: string, lines: readonly OCRDetection[]): string {
+export function fillTableCells(html: string, lines: readonly OcrDetection[]): string {
   'worklet';
-  // Grid size straight from the structure: rows = <tr> count, cols = widest row.
-  const nRows = (html.match(/<tr>/g) ?? []).length;
-  let nCols = 0;
-  const trRe = /<tr>([\s\S]*?)<\/tr>/g;
-  let tr: RegExpExecArray | null;
-  while ((tr = trRe.exec(html)) !== null) {
-    nCols = Math.max(nCols, (tr[1]!.match(/<td/g) ?? []).length);
+  const rowCount = (html.match(/<tr>/g) ?? []).length;
+  let colCount = 0;
+  const rowRegex = /<tr>([\s\S]*?)<\/tr>/g;
+  let row: RegExpExecArray | null;
+  while ((row = rowRegex.exec(html)) !== null) {
+    colCount = Math.max(colCount, (row[1]!.match(/<td/g) ?? []).length);
   }
-  // No grid or nothing to place: plain document-order fill of the skeleton.
-  if (nRows === 0 || nCols === 0 || lines.length === 0) {
+  // No grid, or nothing to place: fill the skeleton cells in document order.
+  if (rowCount === 0 || colCount === 0 || lines.length === 0) {
     let i = 0;
-    return html.replace(/<td([^>]*)><\/td>/g, (_m, attrs) => {
+    return html.replace(/<td([^>]*)><\/td>/g, (_match, attrs) => {
       const text = i < lines.length ? lines[i]!.text : '';
       i++;
       return `<td${attrs}>${text}</td>`;
     });
   }
 
-  const cx: number[] = [];
-  const cy: number[] = [];
-  for (const l of lines) {
-    const b = boundingBoxOf(l.quad);
-    cx.push((b.xmin + b.xmax) / 2);
-    cy.push((b.ymin + b.ymax) / 2);
+  const centersX: number[] = [];
+  const centersY: number[] = [];
+  for (const line of lines) {
+    const box = boundingBoxOf(line.quad);
+    centersX.push((box.xmin + box.xmax) / 2);
+    centersY.push((box.ymin + box.ymax) / 2);
   }
-  const rowC = splitIntoK(cy, nRows);
-  const colC = splitIntoK(cx, nCols);
-  const grid: string[][] = Array.from({ length: rowC.length }, () =>
-    new Array<string>(colC.length).fill('')
+  const rowCenters = clusterCentersByGaps(centersY, rowCount);
+  const colCenters = clusterCentersByGaps(centersX, colCount);
+  const grid: string[][] = Array.from({ length: rowCenters.length }, () =>
+    new Array<string>(colCenters.length).fill('')
   );
   for (let i = 0; i < lines.length; i++) {
-    const r = nearest(cy[i]!, rowC);
-    const c = nearest(cx[i]!, colC);
+    const r = nearestIndex(centersY[i]!, rowCenters);
+    const c = nearestIndex(centersX[i]!, colCenters);
     grid[r]![c] = `${grid[r]![c]!} ${lines[i]!.text}`.trim();
   }
   return `<table>${grid
-    .map((row) => `<tr>${row.map((t) => `<td>${t}</td>`).join('')}</tr>`)
+    .map((cells) => `<tr>${cells.map((text) => `<td>${text}</td>`).join('')}</tr>`)
     .join('')}</table>`;
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
new file mode 100644
index 0000000000..2fb3e65ad0
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
@@ -0,0 +1,329 @@
+import type { WorkletRuntime } from 'react-native-worklets';
+
+import { tensor, type Tensor } from '../../../../core/tensor';
+import { loadModel } from '../../../../core/model';
+import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
+import { wrapAsync } from '../../../../core/runtime';
+
+import type { ImageBuffer, ImageFormat } from '../../image';
+import { IMAGENET_NORM } from '../../../../constants';
+import { FORMAT_CHANNELS } from '../../ops/image';
+import { warpByGrid } from '../../ops/image';
+import { createImagePreprocessor } from '../preprocessing';
+
+/**
+ * A detected page orientation: the clockwise rotation (rotate by its negation to
+ * correct) and the classifier confidence in `[0, 1]`.
+ * @category Types
+ */
+export type Orientation = {
+  readonly rotationCW: 0 | 90 | 180 | 270;
+  readonly confidence: number;
+};
+
+/**
+ * A recognized table structure: the `<tr>/<td>` HTML skeleton (empty cells) and
+ * the raw structure token ids it was built from (start/end tokens stripped).
+ * @category Types
+ */
+export type TableStructure = {
+  readonly html: string;
+  readonly tokens: number[];
+};
+
+/**
+ * Configuration for the fused document models — page orientation, geometric
+ * dewarp, and table-structure recognition, all exposed by one model file.
+ * `structureVocab` maps a table token id (array index) to its HTML fragment;
+ * `eosTokenId` ends table decoding and `maxSteps` caps it.
+ * @category Types
+ */
+export type DocumentModelsConfig = {
+  readonly modelPath: string;
+  readonly structureVocab: readonly string[];
+  readonly eosTokenId: number;
+  readonly maxSteps: number;
+};
+
+// A dewarp grid estimated on a page without clear boundaries (e.g. text floating
+// on white) can map most of the output off the source, collapsing the page to
+// near-blank and OCR to zero detections. The dewarp guard compares content before
+// and after: if the warp keeps less than this fraction of the source's activity,
+// it is declined and the original page is kept.
+const DEWARP_MIN_ACTIVITY_RATIO = 0.5;
+const DEWARP_ACTIVITY_STRIDE = 31;
+
+// The variance of one channel, sampled every DEWARP_ACTIVITY_STRIDE pixels — a
+// cheap, polarity-independent proxy for how much content (ink/edges) an image
+// carries; a blank page is ~0. Used by the dewarp degeneracy guard.
+function dewarpActivity(data: Uint8Array, channels: number): number {
+  'worklet';
+  let n = 0;
+  let sum = 0;
+  let sumSq = 0;
+  const step = channels * DEWARP_ACTIVITY_STRIDE;
+  for (let i = 0; i < data.length; i += step) {
+    const v = data[i]!;
+    sum += v;
+    sumSq += v * v;
+    n++;
+  }
+  if (n === 0) {
+    return 0;
+  }
+  const mean = sum / n;
+  return sumSq / n - mean * mean;
+}
+
+// Index of the maximum value in `arr[offset, offset+len)` (single pass, no allocation).
+function argmaxRange(arr: ArrayLike<number>, offset: number, len: number): number {
+  'worklet';
+  let index = 0;
+  let best = arr[offset]!;
+  for (let i = 1; i < len; i++) {
+    const value = arr[offset + i]!;
+    if (value > best) {
+      best = value;
+      index = i;
+    }
+  }
+  return index;
+}
+
+// Assembles the table-structure content tokens into an HTML skeleton, dropping the
+// reserved start/end range.
+function tokensToHtml(
+  tokens: number[],
+  structureVocab: readonly string[],
+  eosTokenId: number
+): string {
+  'worklet';
+  let html = '';
+  for (const t of tokens) {
+    if (t > 0 && t < eosTokenId && t < structureVocab.length) {
+      html += structureVocab[t]!;
+    }
+  }
+  return html;
+}
+
+/**
+ * Creates the document-models runner: page orientation, geometric dewarp (applied
+ * via the predicted sampling grid), and table-structure recognition (autoregressive
+ * decode). One model file, loaded once. Internal to the document pipeline.
+ * @category Typescript API
+ * @param config Model path, table-structure vocabulary, and decode limits.
+ * @param runtime Optional worklet runtime thread.
+ * @returns A promise resolving to the three capabilities plus disposal controls.
+ */
+export async function createDocumentModels(
+  config: DocumentModelsConfig,
+  runtime?: WorkletRuntime
+): Promise<{
+  dispose: () => void;
+  detectOrientation: (page: Tensor, format: ImageFormat) => Promise<Orientation>;
+  detectOrientationWorklet: (page: Tensor, format: ImageFormat) => Orientation;
+  dewarp: (page: Tensor, format: ImageFormat) => Promise<Tensor>;
+  dewarpWorklet: (page: Tensor, format: ImageFormat) => Tensor;
+  recognizeTable: (input: ImageBuffer) => Promise<TableStructure>;
+  recognizeTableWorklet: (input: ImageBuffer) => TableStructure;
+}> {
+  const { modelPath, structureVocab, eosTokenId, maxSteps } = config;
+  const model = await wrapAsync(loadModel, runtime)(modelPath);
+
+  // Contract validation + preprocessor/tensor construction below can throw (a
+  // missing method, shape/vocab mismatch, failed allocation). Everything built
+  // is pushed into `created` as it is created — one by one, so a mid-sequence
+  // failure can't strand its predecessors — and the catch disposes it all: a
+  // bad config must not leak native memory (mirrors createOcr).
+  const created: { dispose: () => void }[] = [];
+  try {
+    // orientation: image -> class logits
+    const oriMeta = validateModelSchema(
+      model,
+      'orientation',
+      [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 'K'])]
+    );
+    // dewarp: image -> sampling grid
+    const dewMeta = validateModelSchema(
+      model,
+      'dewarp',
+      [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 2, 'gH', 'gW'])]
+    );
+    // table: image encoder + autoregressive decode step
+    const encMeta = validateModelSchema(
+      model,
+      'table_encode',
+      [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 'C', 'F'])]
+    );
+    const decMeta = validateModelSchema(
+      model,
+      'table_decode_step',
+      [
+        SymbolicTensor('float32', [1, 'C', 'F']),
+        SymbolicTensor('float32', [1, 'H']),
+        SymbolicTensor('float32', [1, 'V']),
+      ],
+      [SymbolicTensor('float32', [1, 'V']), SymbolicTensor('float32', [1, 'H'])]
+    );
+
+    const oriShape = oriMeta.inputTensorMeta[0]!.shape;
+    const oriOutLen = oriMeta.outputTensorMeta[0]!.shape[1]!;
+    const dewShape = dewMeta.inputTensorMeta[0]!.shape;
+    const gridShape = dewMeta.outputTensorMeta[0]!.shape; // [1,2,gH,gW]
+    const tabShape = encMeta.inputTensorMeta[0]!.shape;
+    const featShape = encMeta.outputTensorMeta[0]!.shape;
+    const hidShape = decMeta.outputTensorMeta[1]!.shape;
+    const probShape = decMeta.outputTensorMeta[0]!.shape;
+    const hidLen = hidShape[1]!;
+    const vocabLen = probShape[1]!;
+
+    if (vocabLen !== structureVocab.length) {
+      throw new Error(
+        `DocumentModels: structure vocab length (${structureVocab.length}) must match the model's token dim (${vocabLen}).`
+      );
+    }
+
+    const orientationPreprocessor = createImagePreprocessor(
+      {
+        resizeMode: 'stretch',
+        interpolation: 'linear',
+        alpha: IMAGENET_NORM.alpha,
+        beta: IMAGENET_NORM.beta,
+      },
+      oriShape
+    );
+    created.push(orientationPreprocessor);
+    const dewarpPreprocessor = createImagePreprocessor(
+      { resizeMode: 'stretch', interpolation: 'linear', alpha: 1 / 255, beta: 0 },
+      dewShape
+    );
+    created.push(dewarpPreprocessor);
+    const tablePreprocessor = createImagePreprocessor(
+      {
+        resizeMode: 'stretch',
+        interpolation: 'linear',
+        alpha: IMAGENET_NORM.alpha,
+        beta: IMAGENET_NORM.beta,
+      },
+      tabShape
+    );
+    created.push(tablePreprocessor);
+
+    const tOri = tensor('float32', oriMeta.outputTensorMeta[0]!.shape);
+    created.push(tOri);
+    const tGrid = tensor('float32', gridShape);
+    created.push(tGrid);
+    const tFeatures = tensor('float32', featShape);
+    created.push(tFeatures);
+    const tHidden = tensor('float32', hidShape);
+    created.push(tHidden);
+    const tOnehot = tensor('float32', probShape);
+    created.push(tOnehot);
+    const tProbs = tensor('float32', probShape);
+    created.push(tProbs);
+    const tNewHidden = tensor('float32', hidShape);
+    created.push(tNewHidden);
+
+    const oriBuf = new Float32Array(oriOutLen);
+    const zeroHidden = new Float32Array(hidLen);
+    const zeroVocab = new Float32Array(vocabLen);
+    const onehotBuf = new Float32Array(vocabLen);
+    const probsBuf = new Float32Array(vocabLen);
+
+    const dispose = () => {
+      created.forEach((c) => c.dispose());
+      model.dispose();
+    };
+
+    const detectOrientationWorklet = (page: Tensor, format: ImageFormat): Orientation => {
+      'worklet';
+      const tInput = orientationPreprocessor.processTensor(page, format);
+      model.execute('orientation', [tInput], [tOri]);
+      tOri.getData(oriBuf);
+      const cls = argmaxRange(oriBuf, 0, oriOutLen);
+      const best = oriBuf[cls]!;
+      let sumExp = 0;
+      for (let i = 0; i < oriOutLen; i++) {
+        sumExp += Math.exp(oriBuf[i]! - best);
+      }
+      const rotationCW = ((cls % 4) * 90) as 0 | 90 | 180 | 270;
+      const confidence = 1 / sumExp;
+      return { rotationCW, confidence };
+    };
+
+    // Dewarps the full-res page tensor in place: estimate the sampling field, apply
+    // it natively (cv::remap). Returns the dewarped tensor, or the input `page`
+    // unchanged when the warp is declined (caller owns whichever is returned).
+    const dewarpWorklet = (page: Tensor, format: ImageFormat): Tensor => {
+      'worklet';
+      const tInput = dewarpPreprocessor.processTensor(page, format);
+      model.execute('dewarp', [tInput], [tGrid]);
+      const h = page.shape[0]!;
+      const w = page.shape[1]!;
+      const ch = FORMAT_CHANNELS[format];
+      const tDst = tensor('uint8', [h, w, ch]);
+      try {
+        warpByGrid(page, tGrid, tDst);
+        const out = new Uint8Array(w * h * ch);
+        const src = new Uint8Array(w * h * ch);
+        tDst.getData(out);
+        page.getData(src);
+        // Degenerate-warp guard: a grid lacking page boundaries can push content
+        // off-canvas, leaving a near-blank page. If the dewarp collapsed the image's
+        // activity, decline it and keep the original (better an un-dewarped read than
+        // zero detections).
+        if (dewarpActivity(out, ch) < DEWARP_MIN_ACTIVITY_RATIO * dewarpActivity(src, ch)) {
+          tDst.dispose();
+          return page;
+        }
+        return tDst;
+      } catch (e) {
+        // On failure the caller can't see tDst to free it (success path returns it),
+        // so release it here before propagating.
+        tDst.dispose();
+        throw e;
+      }
+    };
+
+    const recognizeTableWorklet = (input: ImageBuffer): TableStructure => {
+      'worklet';
+      const tInput = tablePreprocessor.process(input);
+      model.execute('table_encode', [tInput], [tFeatures]);
+      tHidden.setData(zeroHidden);
+      tOnehot.setData(zeroVocab);
+      const tokens: number[] = [];
+      for (let step = 0; step < maxSteps; step++) {
+        model.execute('table_decode_step', [tFeatures, tHidden, tOnehot], [tProbs, tNewHidden]);
+        tProbs.getData(probsBuf);
+        const tok = argmaxRange(probsBuf, 0, vocabLen);
+        tokens.push(tok);
+        if (tok === eosTokenId) {
+          break;
+        }
+        tNewHidden.copyTo(tHidden);
+        onehotBuf.fill(0);
+        onehotBuf[tok] = 1;
+        tOnehot.setData(onehotBuf);
+      }
+      return { html: tokensToHtml(tokens, structureVocab, eosTokenId), tokens };
+    };
+
+    return {
+      dispose,
+      detectOrientation: wrapAsync(detectOrientationWorklet, runtime),
+      detectOrientationWorklet,
+      dewarp: wrapAsync(dewarpWorklet, runtime),
+      dewarpWorklet,
+      recognizeTable: wrapAsync(recognizeTableWorklet, runtime),
+      recognizeTableWorklet,
+    };
+  } catch (e) {
+    created.forEach((c) => c.dispose());
+    model.dispose();
+    throw e;
+  }
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
deleted file mode 100644
index fc467231af..0000000000
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrHelpers.ts
+++ /dev/null
@@ -1,542 +0,0 @@
-import type { Quad } from '../../ops/textBoxes';
-import { scalePoint, type Point } from '../../ops/points';
-
-/**
- * Worklet-safe millisecond clock for in-pipeline profiling. Prefers
- * `performance.now()` (sub-ms) and falls back to `Date.now()`.
- * @category Typescript API
- * @returns The current time in milliseconds.
- */
-export function nowMs(): number {
-  'worklet';
-  const perf = (globalThis as { performance?: { now?: () => number } }).performance;
-  return perf?.now ? perf.now() : Date.now();
-}
-
-/**
- * The static input-size buckets a bucketed OCR model exposes. Each PTE ships
- * per-size methods `detect_<S>` (square `S×S` detector input) and
- * `recognize_<W>` (fixed recognizer height, width `W`). The client snaps each
- * input to the closest bucket and calls the matching method. Mirrors each
- * model's `config.json` `buckets` (today hardcoded in `models.ts`). Both lists
- * must be ascending.
- * @category Types
- */
-export type Buckets = {
-  /** Square detector input sides (e.g. `[640, 960, 1280]`), ascending. */
-  readonly detect: readonly number[];
-  /** Recognizer input widths (e.g. `[160, 320, 480, 640]`), ascending. */
-  readonly recognize: readonly number[];
-};
-
-/**
- * Snaps an image's longest side *up* to the smallest detector bucket that doesn't
- * downscale it (clamped to the largest bucket), and returns that square side.
- * Rounding up, not nearest, is deliberate: downscaling below the detector's
- * resolution drops text detail and garbles dense lines.
- * @category Typescript API
- * @param imgW Original image width.
- * @param imgH Original image height.
- * @param buckets The ascending detector side buckets.
- * @returns The selected square side `S` (call `detect_${S}`).
- */
-export function snapDetectBucket(imgW: number, imgH: number, buckets: readonly number[]): number {
-  'worklet';
-  const longest = Math.max(imgW, imgH);
-  for (const s of buckets) {
-    if (s >= longest) {
-      return s;
-    }
-  }
-  return buckets[buckets.length - 1]!;
-}
-
-/**
- * Snaps a desired recognizer content width up to the smallest bucket that fits
- * it (avoiding horizontal glyph compression), clamped to the largest bucket.
- * @category Typescript API
- * @param desiredW The crop/feature content width at the recognizer height.
- * @param buckets The ascending recognizer width buckets.
- * @returns The selected width `W` (call `recognize_${W}`).
- */
-export function snapRecognizeBucket(desiredW: number, buckets: readonly number[]): number {
-  'worklet';
-  for (const w of buckets) {
-    if (w >= desiredW) {
-      return w;
-    }
-  }
-  return buckets[buckets.length - 1]!;
-}
-
-/**
- * Maps a detector-space quad back into original image pixel coordinates
- * (reversing the aspect-preserving letterbox), clamping to the image bounds.
- * @category Typescript API
- * @param quad The detector-space quad.
- * @param detW The detector input width the quad is expressed in.
- * @param detH The detector input height the quad is expressed in.
- * @param origW Original image width.
- * @param origH Original image height.
- * @returns The four quad corners in original image pixels.
- */
-export function mapQuadToImage(
-  quad: Quad,
-  detW: number,
-  detH: number,
-  origW: number,
-  origH: number
-): Point[] {
-  'worklet';
-  return quad.points.map((p) => {
-    const m = scalePoint(p, {
-      from: { width: detW, height: detH },
-      to: { width: origW, height: origH },
-      resizeMode: 'letterbox',
-    });
-    return { x: Math.max(0, Math.min(m.x, origW)), y: Math.max(0, Math.min(m.y, origH)) };
-  });
-}
-
-/**
- * Orders four points as TL, TR, BR, BL using x±y extremes (pyimagesearch
- * convention).
- * @category Typescript API
- * @param pts The four unordered quad corners.
- * @returns The corners ordered TL, TR, BR, BL.
- */
-export function orderQuad(pts: readonly Point[]): Point[] {
-  'worklet';
-  // The x±y-extreme ordering below assumes exactly four corners; for any other
-  // count return a copy unchanged rather than indexing out of range.
-  if (pts.length !== 4) {
-    return [...pts];
-  }
-  let tl = 0;
-  let tr = 0;
-  let br = 0;
-  let bl = 0;
-  let minSum = pts[0]!.x + pts[0]!.y;
-  let maxSum = minSum;
-  let minDiff = pts[0]!.y - pts[0]!.x;
-  let maxDiff = minDiff;
-  for (let i = 1; i < 4; i++) {
-    const s = pts[i]!.x + pts[i]!.y;
-    const d = pts[i]!.y - pts[i]!.x;
-    if (s < minSum) {
-      minSum = s;
-      tl = i;
-    }
-    if (s > maxSum) {
-      maxSum = s;
-      br = i;
-    }
-    if (d < minDiff) {
-      minDiff = d;
-      tr = i;
-    }
-    if (d > maxDiff) {
-      maxDiff = d;
-      bl = i;
-    }
-  }
-  return [pts[tl]!, pts[tr]!, pts[br]!, pts[bl]!];
-}
-
-const distance = (a: Point, b: Point): number => {
-  'worklet';
-  return Math.hypot(b.x - a.x, b.y - a.y);
-};
-
-// Linear interpolation between two points (t in [0,1]). Module-level so the
-// worklet plugin captures it; used by splitTallQuad to cut a quad into bands.
-const lerp = (a: Point, b: Point, t: number): Point => {
-  'worklet';
-  return { x: a.x + (b.x - a.x) * t, y: a.y + (b.y - a.y) * t };
-};
-
-/**
- * The natural width/height (in pixels) of an ordered TL,TR,BR,BL quad.
- * @category Typescript API
- * @param ordered The quad corners ordered TL, TR, BR, BL.
- * @returns The quad's width and height in pixels.
- */
-export function quadSize(ordered: readonly Point[]): { width: number; height: number } {
-  'worklet';
-  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
-  const width = Math.max(distance(tl, tr), distance(bl, br));
-  const height = Math.max(distance(tl, bl), distance(tr, br));
-  return { width, height };
-}
-
-// Reading-order tuning. A vertical gutter crossed by no box must be at least this
-// fraction of the content width to separate columns; two boxes share a text line
-// when their vertical extents overlap by at least this fraction of the shorter
-// box's height. Stable across models — not worth exposing.
-const READING_COL_GAP_FRAC = 0.06;
-const READING_ROW_OVERLAP_FRAC = 0.3;
-
-type ReadingBox = { xmin: number; ymin: number; xmax: number; ymax: number };
-
-// Axis-aligned bounds of a quad. Defined before readingOrderIndices so the worklet
-// plugin captures it (a referenced worklet must precede its caller in source order).
-function bboxOfQuad(quad: readonly Point[]): ReadingBox {
-  'worklet';
-  if (quad.length === 0) {
-    return { xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
-  }
-  let xmin = Infinity;
-  let ymin = Infinity;
-  let xmax = -Infinity;
-  let ymax = -Infinity;
-  for (const p of quad) {
-    if (p.x < xmin) xmin = p.x;
-    if (p.y < ymin) ymin = p.y;
-    if (p.x > xmax) xmax = p.x;
-    if (p.y > ymax) ymax = p.y;
-  }
-  return { xmin, ymin, xmax, ymax };
-}
-
-// Horizontal sort key for a box: edge sum = 2× center-x (monotonic in center,
-// avoids the divide). Module-level so it isn't re-created per readingOrderIndices.
-function xSum(b: ReadingBox): number {
-  'worklet';
-  return b.xmin + b.xmax;
-}
-
-/**
- * Returns the indices of `quads` in human reading order. Detects vertical column
- * gutters with an x-coverage sweep — a band crossed by no box and wider than
- * {@link READING_COL_GAP_FRAC} of the content width splits columns (within-line
- * word gaps don't, because other lines cover that x); within each column it groups
- * boxes into lines by vertical overlap, orders lines top-to-bottom and the boxes in
- * a line left-to-right, and reads columns left-to-right. Single-line / single-column
- * inputs collapse to the obvious order. The detector emits boxes in an arbitrary
- * order, so callers reorder their detections/lines through this.
- * @category Typescript API
- * @param quads The detected text quads (any corner order).
- * @returns Indices into `quads`, in reading order.
- */
-export function readingOrderIndices(quads: readonly (readonly Point[])[]): number[] {
-  'worklet';
-  const n = quads.length;
-  if (n <= 1) {
-    return n === 1 ? [0] : [];
-  }
-  const boxes = quads.map((q) => bboxOfQuad(q));
-  let X0 = Infinity;
-  let X1 = -Infinity;
-  for (const b of boxes) {
-    if (b.xmin < X0) X0 = b.xmin;
-    if (b.xmax > X1) X1 = b.xmax;
-  }
-  const minGap = READING_COL_GAP_FRAC * Math.max(1, X1 - X0);
-
-  // Column detection: sweep box x-edges; an internal span where coverage drops to
-  // zero (no box at that x across any line) wider than minGap is a column gutter.
-  const edges: { x: number; d: number }[] = [];
-  for (const b of boxes) {
-    edges.push({ x: b.xmin, d: 1 });
-    edges.push({ x: b.xmax, d: -1 });
-  }
-  // At equal x, open (+1) before close (-1) so touching boxes don't open a gutter.
-  edges.sort((a, b) => a.x - b.x || b.d - a.d);
-  const cuts: number[] = [];
-  let active = 0;
-  let gutterStart = 0;
-  for (const e of edges) {
-    const before = active;
-    active += e.d;
-    if (before > 0 && active === 0) {
-      gutterStart = e.x;
-    } else if (before === 0 && active > 0 && e.x - gutterStart >= minGap) {
-      cuts.push((gutterStart + e.x) / 2);
-    }
-  }
-
-  // Bucket boxes into columns by center-x relative to the (ascending) cut lines.
-  const numCols = cuts.length + 1;
-  const columns: number[][] = [];
-  for (let c = 0; c < numCols; c++) {
-    columns.push([]);
-  }
-  for (let i = 0; i < n; i++) {
-    const cx = (boxes[i]!.xmin + boxes[i]!.xmax) / 2;
-    let c = 0;
-    for (const cut of cuts) {
-      if (cx > cut) c++;
-    }
-    columns[c]!.push(i);
-  }
-
-  // Within each column: group boxes into lines by vertical overlap, order lines
-  // top-to-bottom and boxes within a line left-to-right.
-  const out: number[] = [];
-  for (const col of columns) {
-    col.sort((a, b) => boxes[a]!.ymin - boxes[b]!.ymin);
-    const rows: { items: number[]; ymin: number; ymax: number }[] = [];
-    for (const i of col) {
-      const b = boxes[i]!;
-      let placed = false;
-      for (const row of rows) {
-        const overlap = Math.min(row.ymax, b.ymax) - Math.max(row.ymin, b.ymin);
-        const minH = Math.min(row.ymax - row.ymin, b.ymax - b.ymin);
-        if (overlap >= READING_ROW_OVERLAP_FRAC * Math.max(1, minH)) {
-          row.items.push(i);
-          row.ymin = Math.min(row.ymin, b.ymin);
-          row.ymax = Math.max(row.ymax, b.ymax);
-          placed = true;
-          break;
-        }
-      }
-      if (!placed) {
-        rows.push({ items: [i], ymin: b.ymin, ymax: b.ymax });
-      }
-    }
-    rows.sort((a, b) => a.ymin - b.ymin);
-    // Sort within a line by horizontal position (edge sum = 2× center; monotonic).
-    for (const row of rows) {
-      row.items.sort((a, b) => xSum(boxes[a]!) - xSum(boxes[b]!));
-      for (const i of row.items) {
-        out.push(i);
-      }
-    }
-  }
-  return out;
-}
-
-// Vertical-text grouping tuning. A box wider than this × its height is a
-// horizontal line, never a stacked-column glyph. A box joins a column when its
-// x-span overlaps the column's by VERTICAL_X_OVERLAP of the narrower width and the
-// y-gap is within VERTICAL_Y_GAP × its height (loose — signage spacing varies).
-const VERTICAL_GLYPH_ASPECT = 1.6;
-const VERTICAL_X_OVERLAP = 0.25;
-const VERTICAL_Y_GAP = 2.5;
-
-/**
- * Divides an ordered TL,TR,BR,BL quad into `parts` equal vertical bands (each a
- * TL,TR,BR,BL quad), top to bottom. Recovers the individual upright letters of a
- * stacked column from a box the detector merged (DBNet emits one box per text
- * region, not per glyph, so stacked letters arrive fused). `parts <= 1` returns
- * the quad unchanged.
- * @category Typescript API
- * @param ordered The quad corners ordered TL, TR, BR, BL.
- * @param parts The number of equal vertical bands to split into.
- * @returns The bands as ordered TL,TR,BR,BL quads, top to bottom.
- */
-export function splitTallQuad(ordered: readonly Point[], parts: number): Point[][] {
-  'worklet';
-  if (parts <= 1) {
-    return [ordered as Point[]];
-  }
-  const [tl, tr, br, bl] = ordered as [Point, Point, Point, Point];
-  const out: Point[][] = [];
-  for (let i = 0; i < parts; i++) {
-    const t0 = i / parts;
-    const t1 = (i + 1) / parts;
-    // Left edge runs tl->bl, right edge tr->br; take the band between t0 and t1.
-    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
-  }
-  return out;
-}
-
-/**
- * The axis-aligned bounding quad (TL,TR,BR,BL) enclosing a set of quads. Returns a
- * zero quad for empty input.
- * @category Typescript API
- * @param quads The quads to enclose.
- * @returns The four corners of the enclosing box, ordered TL, TR, BR, BL.
- */
-export function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
-  'worklet';
-  const all: Point[] = [];
-  for (const q of quads) {
-    for (const p of q) {
-      all.push(p);
-    }
-  }
-  const { xmin, ymin, xmax, ymax } = bboxOfQuad(all);
-  return [
-    { x: xmin, y: ymin },
-    { x: xmax, y: ymin },
-    { x: xmax, y: ymax },
-    { x: xmin, y: ymax },
-  ];
-}
-
-/**
- * Clusters glyph-like, x-aligned, stacked boxes into vertical columns; wide lines
- * and isolated boxes come back as `singles` to read normally. So the vertical pass
- * ADDS column reading without disturbing horizontal reads.
- * @category Typescript API
- * @param quads The detected text quads (ordered TL,TR,BR,BL).
- * @returns The detected `columns` (each a top-to-bottom list of quads) and the
- * leftover `singles` (horizontal lines / isolated boxes) to read normally.
- */
-export function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
-  columns: Point[][][];
-  singles: Point[][];
-} {
-  'worklet';
-  type B = {
-    q: Point[];
-    xmin: number;
-    xmax: number;
-    ymin: number;
-    ymax: number;
-    w: number;
-    h: number;
-  };
-  const candidates: B[] = [];
-  const singles: Point[][] = [];
-  for (const q of quads) {
-    const { xmin, ymin, xmax, ymax } = bboxOfQuad(q);
-    const w = xmax - xmin;
-    const h = ymax - ymin;
-    if (w > h * VERTICAL_GLYPH_ASPECT) {
-      singles.push(q as Point[]); // a horizontal line — read normally
-    } else {
-      candidates.push({ q: q as Point[], xmin, xmax, ymin, ymax, w, h });
-    }
-  }
-  // Top -> bottom, growing each column from its current bottom box. Alignment is
-  // checked against the column's accumulated x-range (not just the last box), so a
-  // narrow glyph like `I` between wider ones doesn't break the run.
-  candidates.sort((a, b) => a.ymin - b.ymin);
-  type Col = { boxes: B[]; xmin: number; xmax: number; bottom: number };
-  const cols: Col[] = [];
-  for (const b of candidates) {
-    let placed = false;
-    for (const col of cols) {
-      const overlap = Math.min(b.xmax, col.xmax) - Math.max(b.xmin, col.xmin);
-      const aligned = overlap > VERTICAL_X_OVERLAP * Math.min(b.w, col.xmax - col.xmin);
-      const gap = b.ymin - col.bottom;
-      if (aligned && gap < VERTICAL_Y_GAP * b.h && gap > -0.5 * b.h) {
-        col.boxes.push(b);
-        col.xmin = Math.min(col.xmin, b.xmin);
-        col.xmax = Math.max(col.xmax, b.xmax);
-        col.bottom = b.ymax;
-        placed = true;
-        break;
-      }
-    }
-    if (!placed) {
-      cols.push({ boxes: [b], xmin: b.xmin, xmax: b.xmax, bottom: b.ymax });
-    }
-  }
-  const columns: Point[][][] = [];
-  for (const col of cols) {
-    if (col.boxes.length >= 2) {
-      columns.push(col.boxes.map((b) => b.q)); // already top -> bottom
-    } else {
-      singles.push(col.boxes[0]!.q);
-    }
-  }
-  return { columns, singles };
-}
-
-/**
- * Resolves the content width (px) of a recognizer crop: the quad resized to the
- * recognizer height keeping aspect, clamped to the model's bucket width.
- * @category Typescript API
- * @param quadW The quad's natural width.
- * @param quadH The quad's natural height.
- * @param recHeight The recognizer input height.
- * @param bucketWidth The recognizer input (canvas) width.
- * @returns The clamped content width in pixels.
- */
-export function contentWidthFor(
-  quadW: number,
-  quadH: number,
-  recHeight: number,
-  bucketWidth: number
-): number {
-  'worklet';
-  const w = Math.round((recHeight * quadW) / Math.max(1, quadH));
-  return Math.max(1, Math.min(w, bucketWidth));
-}
-
-/**
- * Flattens an ordered TL,TR,BR,BL quad into the 8-number `[x0,y0,..,x3,y3]`
- * array expected by `warpQuad`.
- * @category Typescript API
- * @param corners The four quad corners (TL, TR, BR, BL).
- * @returns The eight coordinates `[x0,y0,x1,y1,x2,y2,x3,y3]`.
- */
-export function flattenQuad(corners: readonly Point[]): number[] {
-  'worklet';
-  // prettier-ignore
-  return [
-    corners[0]!.x, corners[0]!.y, corners[1]!.x, corners[1]!.y,
-    corners[2]!.x, corners[2]!.y, corners[3]!.x, corners[3]!.y,
-  ];
-}
-
-/**
- * Builds a CTC charset array from a charset string. The first `numSpecials`
- * indices are reserved special tokens (the CTC blank for CRNN/SVTR); subsequent
- * entries are the codepoints of `charset` (UTF-8 aware via `Array.from`), so
- * `charset[id]` decodes logit index `id`.
- * @category Typescript API
- * @param charset The ordered character set string.
- * @param numSpecials Number of reserved low indices (default 1 = CTC blank).
- * @returns The charset array with `numSpecials` placeholders at the front.
- */
-export function buildCharset(charset: string | readonly string[], numSpecials = 1): string[] {
-  'worklet';
-  const specials: string[] = [];
-  for (let i = 0; i < numSpecials; i++) {
-    specials.push(`[special${i}]`);
-  }
-  // A string charset is one codepoint per index; an array charset is taken
-  // verbatim (its entries may be multi-codepoint, e.g. ligatures or combining
-  // marks, which `Array.from` would otherwise split).
-  const chars = typeof charset === 'string' ? Array.from(charset) : charset;
-  return [...specials, ...chars];
-}
-
-/**
- * Greedy CTC decode: skip the reserved special/blank indices (`< numSpecials`)
- * and consecutive repeats, mapping indices to charset entries.
- * @category Typescript API
- * @param indices The per-timestep argmax indices.
- * @param charset The charset array (specials/blank at the front).
- * @param numSpecials Number of reserved low indices to drop (default 1).
- * @returns The decoded text.
- */
-export function decodeGreedy(indices: number[], charset: string[], numSpecials = 1): string {
-  'worklet';
-  let text = '';
-  let last = -1;
-  for (const idx of indices) {
-    if (idx >= numSpecials && idx !== last && idx < charset.length) {
-      text += charset[idx]!;
-    }
-    last = idx;
-  }
-  return text;
-}
-
-/**
- * CTC confidence: the mean of the non-special per-timestep max probabilities
- * (the unified contract softmaxes both heads, so this is comparable across
- * models). Skips the reserved indices (`< numSpecials`).
- * @category Typescript API
- * @param values The per-timestep max probabilities.
- * @param indices The per-timestep argmax indices.
- * @param numSpecials Number of reserved low indices to skip (default 1).
- * @returns The aggregate confidence in `[0, 1]`.
- */
-export function ctcConfidence(values: number[], indices: number[], numSpecials = 1): number {
-  'worklet';
-  let sum = 0;
-  let count = 0;
-  for (let i = 0; i < indices.length; i++) {
-    if (indices[i]! >= numSpecials) {
-      sum += values[i]!;
-      count++;
-    }
-  }
-  return count === 0 ? 0 : sum / count;
-}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
deleted file mode 100644
index 20745b7ac5..0000000000
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrPipeline.ts
+++ /dev/null
@@ -1,539 +0,0 @@
-// OCR pipeline engine: the per-page / per-box worklet functions and the
-// construction-time builders/validators behind `createOCR` (in ocr.ts). Split
-// out to keep the task file focused on the public API + factory wiring.
-// Everything here is internal — it is NOT re-exported from the package index.
-//
-// Worklet source-order rule: the worklet plugin captures referenced worklets in
-// source order, so a worklet must be defined BEFORE any worklet that calls it
-// (readStackedColumn -> detectQuads/recognizeGlyphStrip; readBoxVertical ->
-// readStackedColumn/recognizeQuad). The non-worklet builders/validators run at
-// construction time on the JS thread and have no such constraint.
-
-import { tensor, type Tensor, type DType } from '../../../../core/tensor';
-import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
-import type { Model } from '../../../../core/model';
-
-import type { ImageFormat } from '../../image';
-import type { Point } from '../../ops/points';
-import {
-  FORMAT_CONVERSION,
-  resize,
-  cvtColor,
-  toChannelsFirst,
-  normalize,
-  type ColorConversionCode,
-} from '../../ops/image';
-import {
-  extractTextBoxes,
-  warpQuad,
-  ctcGreedyDecode,
-  type CraftExtractOptions,
-  type DbnetExtractOptions,
-  type Quad,
-} from '../../ops/textBoxes';
-import {
-  mapQuadToImage,
-  orderQuad,
-  quadSize,
-  contentWidthFor,
-  flattenQuad,
-  decodeGreedy,
-  ctcConfidence,
-  snapDetectBucket,
-  snapRecognizeBucket,
-  splitTallQuad,
-} from './ocrHelpers';
-
-// Detector input is raw RGB scaled to [0,1]; the per-architecture mean/std is
-// baked into the PTE, so the client only divides by 255.
-const DETECTOR_ALPHA = 1 / 255;
-const DETECTOR_BETA = 0;
-
-// Detector box-extraction tuning. These are stable across models and not worth
-// exposing as options — bake them in. (CRAFT = EasyOCR, DBNet = PaddleOCR.)
-const CRAFT_TEXT_THRESHOLD = 0.4;
-const CRAFT_LINK_THRESHOLD = 0.4;
-const CRAFT_LOW_TEXT_THRESHOLD = 0.7;
-const DBNET_BIN_THRESHOLD = 0.3;
-const DBNET_BOX_THRESHOLD = 0.6;
-const DBNET_UNCLIP_RATIO = 1.5;
-const DBNET_MIN_BOX_SIDE = 3;
-const DBNET_MAX_CANDIDATES = 1000;
-// DBNet exports a post-sigmoid prob map, so don't re-apply sigmoid (true only for
-// raw-logit heads).
-const APPLY_SIGMOID = false;
-
-// Stored at construction; CRAFT's `targetHeight` is per-run, so detectQuads adds it.
-export type DetectorExtractConfig = Omit<CraftExtractOptions, 'targetHeight'> | DbnetExtractOptions;
-
-// Per-detect-bucket scratch tensors, allocated once at construction (like RecSet)
-// and reused across the page pass and per-box re-detects. Only the source-resize
-// tensor depends on the run's input channel count, so detectQuads allocates that
-// one per call and reuses the rest from here.
-export type DetSet = {
-  readonly s: number;
-  readonly tColor: Tensor; // [s, s, 3]
-  readonly tCF: Tensor; // [3, s, s]
-  readonly tNorm: Tensor; // [3, s, s]
-  readonly tInput: Tensor; // [1, 3, s, s]
-  // The detector's output tensors (dbnet prob-map / craft heatmap+extras / a
-  // custom arch's raw outputs); the built-in decoder reads tOutputs[0].
-  readonly tOutputs: readonly Tensor[];
-};
-
-// Everything the detector pass needs, bundled so it can run both on the full
-// page and (for vertical text) on a single box crop to find its characters.
-export type DetectContext = {
-  readonly model: Model;
-  readonly detBuckets: readonly number[];
-  readonly numChannels: number;
-  readonly detCode: ColorConversionCode | null;
-  // Built-in decode config (craft/dbnet); undefined when a custom extractor is used.
-  readonly extractOpts?: DetectorExtractConfig;
-  // Custom TS box extractor for detectorKind 'custom'; takes precedence when set.
-  readonly extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[];
-  readonly detSets: ReadonlyMap<number, DetSet>;
-};
-
-// A recognizer width bucket's pre-allocated tensor-set (one per width in the
-// model's `recognize` buckets).
-export type RecSet = {
-  readonly width: number;
-  readonly tCanvas: Tensor;
-  readonly tCF: Tensor;
-  readonly tNorm: Tensor;
-  readonly tInput: Tensor;
-  readonly tLogits: Tensor;
-};
-
-// Recognizer state for reading one quad. The source image is passed per call,
-// so the same context reads both the full page and a per-box crop.
-export type RecContext = {
-  readonly model: Model;
-  readonly recSetByWidth: ReadonlyMap<number, RecSet>;
-  readonly recBuckets: readonly number[];
-  readonly recH: number;
-  readonly charset: string[];
-  // Per-model recognizer normalization / pad (resolved from OCROptions defaults).
-  readonly normAlpha: number | readonly number[];
-  readonly normBeta: number | readonly number[];
-  readonly padValue: number;
-  // Optional custom decode; falls back to greedy CTC when absent.
-  readonly decode?: (
-    logits: Tensor,
-    charset: readonly string[]
-  ) => { readonly text: string; readonly confidence: number };
-};
-
-// State the vertical-text path needs on top of `RecContext`: the detector (for
-// the second, character-level pass) and the page tensors it crops the box from.
-export type VerticalContext = {
-  readonly detCtx: DetectContext;
-  readonly rawPage: Tensor; // full page, native format — cropped per box for re-detection
-  readonly recCode: ColorConversionCode | null; // native format -> recognizer color
-  readonly recC: number;
-  readonly tallCropRatio: number; // height/width ratio above which a box is a column
-  // Per-page budget for the (expensive) stacked-column re-detection pass.
-  readonly redetectBudget: { remaining: number };
-};
-
-// cvtColor code to bring a source image format to RGB (both the detector and the
-// recognizer operate on RGB), or null when it already is RGB.
-export function toRgbCode(format: ImageFormat): ColorConversionCode | null {
-  'worklet';
-  return FORMAT_CONVERSION[format].rgb;
-}
-
-// Detects text boxes in `src` (uint8 [H,W,numChannels], native `format`) and
-// returns quads in `src` pixel space: letterbox into the snapped square bucket,
-// run `detect_<S>`, extract quads. Scratch is the bucket's cached DetSet; only
-// the source-resize tensor (input-channel-dependent) is allocated/freed here.
-export function detectQuads(
-  ctx: DetectContext,
-  src: Tensor,
-  width: number,
-  height: number,
-  charLevel = false
-): Point[][] {
-  'worklet';
-  const detS = snapDetectBucket(width, height, ctx.detBuckets);
-  // snapDetectBucket always returns one of detBuckets, so the set exists.
-  const detSet = ctx.detSets.get(detS)!;
-  // Only the source resize depends on the run's channel count; the rest is cached.
-  const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
-  try {
-    src
-      .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
-      .throughIf(ctx.detCode !== null, cvtColor, detSet.tColor, ctx.detCode!)
-      .through(toChannelsFirst, detSet.tCF)
-      .through(normalize, detSet.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
-      .copyTo(detSet.tInput);
-
-    ctx.model.execute(`detect_${detS}`, [detSet.tInput], [...detSet.tOutputs]);
-    // A custom arch hands its raw outputs to the user extractor; the built-ins
-    // decode the heatmap (tOutputs[0]). CRAFT needs the per-run input height to
-    // restore its half-res boxes; `charLevel` switches to per-glyph boxes.
-    const quads = ctx.extractBoxes
-      ? ctx.extractBoxes(detSet.tOutputs, detS)
-      : extractTextBoxes(
-          detSet.tOutputs[0]!,
-          ctx.extractOpts!.mode === 'craft'
-            ? { ...ctx.extractOpts!, targetHeight: detS, charLevel }
-            : ctx.extractOpts!
-        );
-    return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
-  } finally {
-    tDetResize.dispose();
-  }
-}
-
-// Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap content width to a
-// recognizer bucket, warp -> normalize -> execute -> greedy-CTC decode.
-export function recognizeQuad(
-  ctx: RecContext,
-  src: Tensor,
-  corners: readonly Point[]
-): { text: string; conf: number } {
-  'worklet';
-  const cs = quadSize(corners);
-  const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
-  const desiredW = contentWidthFor(cs.width, cs.height, ctx.recH, maxRec);
-  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
-  // snapRecognizeBucket always returns one of recBuckets, so the set exists.
-  const recSet = ctx.recSetByWidth.get(bucketW)!;
-  const contentWidth = Math.min(desiredW, bucketW);
-  warpQuad(src, recSet.tCanvas, flattenQuad(corners), {
-    contentWidth,
-    align: 'left',
-    padMode: 'constant',
-    padValue: ctx.padValue,
-  });
-  recSet.tCanvas
-    .through(toChannelsFirst, recSet.tCF)
-    .through(normalize, recSet.tNorm, { alpha: ctx.normAlpha, beta: ctx.normBeta })
-    .copyTo(recSet.tInput);
-  ctx.model.execute(`recognize_${bucketW}`, [recSet.tInput], [recSet.tLogits]);
-  // A custom decode (e.g. attention/AR head) takes the raw logits; otherwise
-  // greedy CTC. Both heads emit probabilities (CRNN softmax baked, SVTR pre-softmaxed).
-  if (ctx.decode) {
-    const r = ctx.decode(recSet.tLogits, ctx.charset);
-    return { text: r.text, conf: r.confidence };
-  }
-  const { indices, values } = ctcGreedyDecode(recSet.tLogits, { softmax: false });
-  const text = decodeGreedy(indices, ctx.charset);
-  const conf = ctcConfidence(values, indices);
-  return { text, conf };
-}
-
-// Joins glyph quads (in `src` pixel space, reading order) into one recognizer
-// strip: each glyph is warped upright to the recognizer height and placed side by
-// side directly in the canvas (native `warpQuad` with a per-glyph `offsetX`, so
-// there is no JS pixel assembly), then recognized as a single line. Returns null
-// when nothing usable was assembled.
-//
-// Must be defined BEFORE its callers: the worklet plugin captures referenced
-// worklets in source order, so a forward reference is undefined at run time.
-export function recognizeGlyphStrip(
-  recCtx: RecContext,
-  src: Tensor,
-  glyphs: readonly (readonly Point[])[]
-): { text: string; conf: number } | null {
-  'worklet';
-  const recH = recCtx.recH;
-  const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-  // Pass 1 (geometry only): a box much taller than wide is a merged run of stacked
-  // letters — split it into ~square single-letter cells so each lands in its own
-  // strip slot. Measure each cell's warped width (aspect preserved) to size the strip.
-  const cells: { quad: readonly Point[]; width: number }[] = [];
-  let totalW = 0;
-  for (const glyph of glyphs) {
-    const glyphSize = quadSize(glyph);
-    if (glyphSize.width < 1 || glyphSize.height < 1) {
-      continue;
-    }
-    const parts = Math.max(1, Math.round(glyphSize.height / Math.max(1, glyphSize.width)));
-    for (const cell of splitTallQuad(glyph, parts)) {
-      const cellSize = quadSize(cell);
-      if (cellSize.width < 1 || cellSize.height < 1) {
-        continue;
-      }
-      const width = Math.max(
-        1,
-        Math.min(Math.round((cellSize.width * recH) / cellSize.height), maxRec)
-      );
-      cells.push({ quad: cell, width });
-      totalW += width;
-    }
-  }
-  if (cells.length === 0) {
-    return null;
-  }
-  // Smallest bucket that fits the strip (snap up, no glyph truncated); widest for
-  // very long columns.
-  const bucketW =
-    recCtx.recBuckets.find((w) => w >= totalW) ?? recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
-  const recSet = recCtx.recSetByWidth.get(bucketW)!;
-  // Pass 2: warp each cell straight into the canvas at its x-offset. The first warp
-  // clears + pads the whole canvas; the rest compose in with `clear: false`.
-  let xOff = 0;
-  for (let i = 0; i < cells.length; i++) {
-    if (xOff >= bucketW) {
-      break;
-    }
-    warpQuad(src, recSet.tCanvas, flattenQuad(cells[i]!.quad), {
-      contentWidth: cells[i]!.width,
-      offsetX: xOff,
-      clear: i === 0,
-      padMode: 'constant',
-      padValue: recCtx.padValue,
-    });
-    xOff += cells[i]!.width;
-  }
-  recSet.tCanvas
-    .through(toChannelsFirst, recSet.tCF)
-    .through(normalize, recSet.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
-    .copyTo(recSet.tInput);
-  recCtx.model.execute(`recognize_${bucketW}`, [recSet.tInput], [recSet.tLogits]);
-  if (recCtx.decode) {
-    const decoded = recCtx.decode(recSet.tLogits, recCtx.charset);
-    return decoded.text.length > 0 ? { text: decoded.text, conf: decoded.confidence } : null;
-  }
-  const { indices, values } = ctcGreedyDecode(recSet.tLogits, { softmax: false });
-  const text = decodeGreedy(indices, recCtx.charset);
-  const conf = ctcConfidence(values, indices);
-  return text.length > 0 ? { text, conf } : null;
-}
-
-// Reads a single tall box that packs several stacked glyphs the detector grouped
-// into one box (e.g. a tightly-set container code): crop it upright, re-detect
-// the glyphs (char-level pass), and read them top -> bottom as a joined strip.
-// Returns null — caller falls back to a horizontal read — when the box is tiny,
-// the per-page re-detect budget is spent, or nothing is found.
-export function readStackedColumn(
-  recCtx: RecContext,
-  vctx: VerticalContext,
-  ordered: readonly Point[],
-  size: { width: number; height: number }
-): { text: string; conf: number } | null {
-  'worklet';
-  const boxW = Math.round(size.width);
-  const boxH = Math.round(size.height);
-  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
-    return null;
-  }
-  vctx.redetectBudget.remaining--;
-  const recC = vctx.recC;
-  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
-  const tRecBox = tensor('uint8', [boxH, boxW, recC]);
-  try {
-    // Axis-aligned upright crop of the box, full width (no padding).
-    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
-      contentWidth: boxW,
-      align: 'left',
-      padMode: 'constant',
-      padValue: 0,
-    });
-    // Char-level second pass: per-glyph boxes (no grouping), in crop pixel space.
-    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
-    if (charQuads.length === 0) {
-      return null;
-    }
-    const boxSrc = vctx.recCode !== null ? cvtColor(tBoxRaw, tRecBox, vctx.recCode) : tBoxRaw;
-    // Stack reading order: top -> bottom by each glyph's upper edge.
-    const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
-    return recognizeGlyphStrip(recCtx, boxSrc, glyphs);
-  } finally {
-    tBoxRaw.dispose();
-    tRecBox.dispose();
-  }
-}
-
-// Reads one box that may be an upright stacked column (e.g. a shipping-container
-// code, letters stacked top-to-bottom). A normal-aspect box is read horizontally;
-// a tall box is read as a stacked column, falling back to the horizontal read.
-// `stacked` reports whether the column path produced the read (caller applies the
-// lenient vertical drop-score to those).
-export function readBoxVertical(
-  recCtx: RecContext,
-  vctx: VerticalContext,
-  pageSrc: Tensor,
-  ordered: readonly Point[],
-  size: { width: number; height: number }
-): { text: string; conf: number; stacked: boolean } {
-  'worklet';
-  if (size.height >= size.width * vctx.tallCropRatio) {
-    const stacked = readStackedColumn(recCtx, vctx, ordered, size);
-    if (stacked) {
-      return { ...stacked, stacked: true };
-    }
-  }
-  return { ...recognizeQuad(recCtx, pageSrc, ordered), stacked: false };
-}
-
-// Validates the detector methods against the architecture's input/output spec
-// and returns CRAFT's extra-output channel counts (empty for dbnet/custom). For
-// 'custom' only the shared RGB input contract is enforced — outputs are read from
-// the PTE metadata and handed to the user's extractBoxes. Throws on any mismatch.
-export function validateDetectorSchema(
-  model: Model,
-  detBuckets: readonly number[],
-  detectorKind: 'craft' | 'dbnet' | 'custom',
-  extractBoxes?: (outputs: readonly Tensor[], s: number) => Quad[]
-): number[] {
-  // CRAFT's half-resolution heatmap needs even detect-bucket sides.
-  if (detectorKind === 'craft' && detBuckets.some((s) => s % 2 !== 0)) {
-    throw new Error('OCR: every CRAFT detect bucket side must be even (half-resolution heatmap).');
-  }
-  const detInSpec = [SymbolicTensor('float32', [1, 3, 'H', 'W'])];
-  if (detectorKind === 'custom') {
-    if (!extractBoxes) {
-      throw new Error("OCR: detectorKind 'custom' requires an extractBoxes worklet.");
-    }
-    // Outputs are arbitrary (read from metadata, handed to extractBoxes); only
-    // the shared RGB input contract is enforced. getMethodMeta throws if missing.
-    for (const s of detBuckets) {
-      const inShape = model.getMethodMeta(`detect_${s}`).inputTensorMeta[0]?.shape;
-      if (!inShape || inShape.length !== 4 || inShape[1] !== 3) {
-        throw new Error(`OCR: detect_${s} must take a [1, 3, ${s}, ${s}] RGB input.`);
-      }
-    }
-    return [];
-  }
-  // Validate every detect bucket against the architecture's output spec; keep
-  // the largest bucket's meta for the constant CRAFT extra-output channels.
-  const detOutSpec =
-    detectorKind === 'dbnet'
-      ? [SymbolicTensor('float32', [1, 1, 'H', 'W'], [1, 'H', 'W'], ['H', 'W'])]
-      : [
-          SymbolicTensor('float32', [1, 'H', 'W', 2], ['H', 'W', 2]),
-          SymbolicTensor('float32', [1, 'C', 'fH', 'fW']),
-        ];
-  const detMeta = validateModelSchema(
-    model,
-    `detect_${detBuckets[detBuckets.length - 1]}`,
-    detInSpec,
-    detOutSpec
-  );
-  for (let i = 0; i < detBuckets.length - 1; i++) {
-    validateModelSchema(model, `detect_${detBuckets[i]}`, detInSpec, detOutSpec);
-  }
-  // CRAFT's extra outputs (feature map) at half resolution; keep the channel counts.
-  return detMeta.outputTensorMeta.slice(1).map((t) => t.shape[1]!);
-}
-
-// Built-in box-decode config for a detector kind (custom archs decode in TS via
-// extractBoxes instead, so they get `undefined`).
-export function buildExtractOpts(
-  detectorKind: 'craft' | 'dbnet' | 'custom'
-): DetectorExtractConfig | undefined {
-  if (detectorKind === 'custom') {
-    return undefined;
-  }
-  return detectorKind === 'dbnet'
-    ? {
-        mode: 'dbnet',
-        binThreshold: DBNET_BIN_THRESHOLD,
-        boxThreshold: DBNET_BOX_THRESHOLD,
-        unclipRatio: DBNET_UNCLIP_RATIO,
-        minBoxSide: DBNET_MIN_BOX_SIDE,
-        maxCandidates: DBNET_MAX_CANDIDATES,
-        applySigmoid: APPLY_SIGMOID,
-      }
-    : {
-        mode: 'craft',
-        textThreshold: CRAFT_TEXT_THRESHOLD,
-        linkThreshold: CRAFT_LINK_THRESHOLD,
-        lowTextThreshold: CRAFT_LOW_TEXT_THRESHOLD,
-      };
-}
-
-// Per-detect-bucket output tensor specs (dtype + shape). Custom archs declare
-// arbitrary outputs — read straight from the PTE's method metadata; the built-ins
-// have known heatmap (+ craft extras) shapes. Returns specs only (no tensors); the
-// task factory allocates and owns them (see DetSet in createOCR). Runs at
-// construction.
-export function deriveDetectorOutputs(
-  model: Model,
-  detBuckets: readonly number[],
-  detectorKind: 'craft' | 'dbnet' | 'custom',
-  detExtraChannels: readonly number[]
-): { s: number; outputs: { dtype: DType; shape: number[] }[] }[] {
-  return detBuckets.map((s) => {
-    const heat = s / 2;
-    if (detectorKind === 'custom') {
-      const outputs = model
-        .getMethodMeta(`detect_${s}`)
-        .outputTensorMeta.map((m) => ({ dtype: m.dtype, shape: m.shape }));
-      return { s, outputs };
-    }
-    const heatmap: { dtype: DType; shape: number[] } =
-      detectorKind === 'dbnet'
-        ? { dtype: 'float32', shape: [1, 1, s, s] }
-        : { dtype: 'float32', shape: [1, heat, heat, 2] };
-    const outputs = [
-      heatmap,
-      ...detExtraChannels.map((c) => ({ dtype: 'float32' as DType, shape: [1, c, heat, heat] })),
-    ];
-    return { s, outputs };
-  });
-}
-
-// Validates each `recognize_<W>` method and derives the recognizer contract:
-// the constant channel/height/vocab (from the first bucket) plus each bucket's
-// input/output shapes. Returns specs only (no tensors); the task factory allocates
-// and owns the RecSet tensors. Runs at construction.
-export function deriveRecognizerContract(
-  model: Model,
-  recBuckets: readonly number[]
-): {
-  recC: number;
-  recH: number;
-  vocabSize: number;
-  buckets: { width: number; inShape: number[]; outShape: number[] }[];
-} {
-  let recC = 0;
-  let recH = 0;
-  let vocabSize = 0;
-  const buckets = recBuckets.map((w, i) => {
-    const m = validateModelSchema(
-      model,
-      `recognize_${w}`,
-      [SymbolicTensor('float32', [1, 'C', 'H', 'W'])],
-      [SymbolicTensor('float32', [1, 'T', 'V'])]
-    );
-    const inShape = m.inputTensorMeta[0]!.shape;
-    const outShape = m.outputTensorMeta[0]!.shape;
-    if (i === 0) {
-      // Channels/height/vocab are constant across the width buckets.
-      recC = inShape[1]!;
-      recH = inShape[2]!;
-      vocabSize = outShape[2]!;
-    }
-    return { width: w, inShape, outShape };
-  });
-  return { recC, recH, vocabSize, buckets };
-}
-
-// Frees a detector scratch-set's tensors (input prep + per-bucket outputs).
-export function disposeDetSets(detSets: readonly DetSet[]): void {
-  for (const d of detSets) {
-    d.tColor.dispose();
-    d.tCF.dispose();
-    d.tNorm.dispose();
-    d.tInput.dispose();
-    for (const t of d.tOutputs) {
-      t.dispose();
-    }
-  }
-}
-
-// Frees a recognizer scratch-set's tensors.
-export function disposeRecSets(recSets: readonly RecSet[]): void {
-  for (const s of recSets) {
-    s.tCanvas.dispose();
-    s.tCF.dispose();
-    s.tNorm.dispose();
-    s.tInput.dispose();
-    s.tLogits.dispose();
-  }
-}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
new file mode 100644
index 0000000000..4059e311a1
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
@@ -0,0 +1,327 @@
+import type { Point } from '../../ops/points';
+import { boundsOfPoints } from '../../ops/quad';
+
+/**
+ * The static input-size buckets a bucketed OCR model exposes. Each model ships
+ * per-size detect (`detect_<S>`, square `S×S` input) and recognize
+ * (`recognize_<W>`, fixed height, width `W`) methods; the pipeline snaps each
+ * input to the nearest bucket and calls the matching method. Both lists must be
+ * ascending.
+ * @category Types
+ */
+export type Buckets = {
+  readonly detect: readonly number[];
+  readonly recognize: readonly number[];
+};
+
+// A size within this fraction above the next-lower bucket snaps DOWN to it rather
+// than up, so a marginal overflow (e.g. 641 px against a 640 bucket) doesn't jump
+// to the next, much larger and slower, bucket. Kept small: a larger downscale
+// loses detail the model was trained to see.
+const BUCKET_SNAP_TOLERANCE = 0.1;
+
+// Selects the smallest bucket that fits `size`, but snaps down to the next-lower
+// bucket when `size` exceeds it by no more than BUCKET_SNAP_TOLERANCE. Clamps to
+// the largest bucket for oversized inputs.
+function snapBucket(size: number, buckets: readonly number[]): number {
+  'worklet';
+  for (let i = 0; i < buckets.length; i++) {
+    if (buckets[i]! >= size) {
+      const lower = i > 0 ? buckets[i - 1]! : 0;
+      return lower > 0 && size <= lower * (1 + BUCKET_SNAP_TOLERANCE) ? lower : buckets[i]!;
+    }
+  }
+  return buckets[buckets.length - 1]!;
+}
+
+/**
+ * Selects the detector bucket for an image from its longest side.
+ * @category Typescript API
+ * @param imageWidth The image width in pixels.
+ * @param imageHeight The image height in pixels.
+ * @param buckets The ascending detector side buckets.
+ * @returns The selected square side `S` (invoke `detect_<S>`).
+ */
+export function snapDetectBucket(
+  imageWidth: number,
+  imageHeight: number,
+  buckets: readonly number[]
+): number {
+  'worklet';
+  return snapBucket(Math.max(imageWidth, imageHeight), buckets);
+}
+
+/**
+ * Selects the recognizer width bucket for a desired crop content width.
+ * @category Typescript API
+ * @param desiredWidth The crop content width at the recognizer height.
+ * @param buckets The ascending recognizer width buckets.
+ * @returns The selected width `W` (invoke `recognize_<W>`).
+ */
+export function snapRecognizeBucket(desiredWidth: number, buckets: readonly number[]): number {
+  'worklet';
+  return snapBucket(desiredWidth, buckets);
+}
+
+/**
+ * Computes the content width (px) of a recognizer crop: the region resized to the
+ * recognizer height keeping its aspect ratio, clamped to the bucket width.
+ * @category Typescript API
+ * @param regionWidth The region's natural width.
+ * @param regionHeight The region's natural height.
+ * @param recognizerHeight The recognizer input height.
+ * @param bucketWidth The recognizer input (canvas) width.
+ * @returns The clamped content width in pixels.
+ */
+export function contentWidthFor(
+  regionWidth: number,
+  regionHeight: number,
+  recognizerHeight: number,
+  bucketWidth: number
+): number {
+  'worklet';
+  const width = Math.round((recognizerHeight * regionWidth) / Math.max(1, regionHeight));
+  return Math.max(1, Math.min(width, bucketWidth));
+}
+
+// A gutter must be at least this fraction of the content width to split columns;
+// two boxes share a line when their vertical extents overlap by at least this
+// fraction of the shorter box's height.
+const COLUMN_GAP_FRACTION = 0.06;
+const LINE_OVERLAP_FRACTION = 0.3;
+
+// Returns the indices of `quads` in human reading order. Column gutters are found
+// by an x-coverage sweep (a band no box crosses, wider than COLUMN_GAP_FRACTION of
+// the content width, splits columns); within each column boxes are grouped into
+// lines by vertical overlap, lines ordered top-to-bottom, boxes within a line
+// left-to-right, and columns read left-to-right.
+function readingOrder(quads: readonly (readonly Point[])[]): number[] {
+  'worklet';
+  const count = quads.length;
+  if (count <= 1) {
+    return count === 1 ? [0] : [];
+  }
+  const boxes = quads.map((q) => boundsOfPoints(q));
+  let minX = Infinity;
+  let maxX = -Infinity;
+  for (const box of boxes) {
+    if (box.xmin < minX) minX = box.xmin;
+    if (box.xmax > maxX) maxX = box.xmax;
+  }
+  const minGap = COLUMN_GAP_FRACTION * Math.max(1, maxX - minX);
+
+  // Sweep the box x-edges; an interior span with zero coverage wider than minGap
+  // is a column gutter, cut at its midpoint.
+  const edges: { x: number; delta: number }[] = [];
+  for (const box of boxes) {
+    edges.push({ x: box.xmin, delta: 1 });
+    edges.push({ x: box.xmax, delta: -1 });
+  }
+  // At equal x, open (+1) before close (-1) so touching boxes don't open a gutter.
+  edges.sort((a, b) => a.x - b.x || b.delta - a.delta);
+  const cuts: number[] = [];
+  let coverage = 0;
+  let gutterStart = 0;
+  for (const edge of edges) {
+    const before = coverage;
+    coverage += edge.delta;
+    if (before > 0 && coverage === 0) {
+      gutterStart = edge.x;
+    } else if (before === 0 && coverage > 0 && edge.x - gutterStart >= minGap) {
+      cuts.push((gutterStart + edge.x) / 2);
+    }
+  }
+
+  // Assign each box to a column by its center-x relative to the (ascending) cuts.
+  const columns: number[][] = Array.from({ length: cuts.length + 1 }, () => []);
+  for (let i = 0; i < count; i++) {
+    const centerX = (boxes[i]!.xmin + boxes[i]!.xmax) / 2;
+    let column = 0;
+    for (const cut of cuts) {
+      if (centerX > cut) column++;
+    }
+    columns[column]!.push(i);
+  }
+
+  const order: number[] = [];
+  for (const column of columns) {
+    column.sort((a, b) => boxes[a]!.ymin - boxes[b]!.ymin);
+    const lines: { items: number[]; ymin: number; ymax: number }[] = [];
+    for (const i of column) {
+      const box = boxes[i]!;
+      let placed = false;
+      for (const line of lines) {
+        const overlap = Math.min(line.ymax, box.ymax) - Math.max(line.ymin, box.ymin);
+        const minHeight = Math.min(line.ymax - line.ymin, box.ymax - box.ymin);
+        if (overlap >= LINE_OVERLAP_FRACTION * Math.max(1, minHeight)) {
+          line.items.push(i);
+          line.ymin = Math.min(line.ymin, box.ymin);
+          line.ymax = Math.max(line.ymax, box.ymax);
+          placed = true;
+          break;
+        }
+      }
+      if (!placed) {
+        lines.push({ items: [i], ymin: box.ymin, ymax: box.ymax });
+      }
+    }
+    lines.sort((a, b) => a.ymin - b.ymin);
+    for (const line of lines) {
+      line.items.sort(
+        (a, b) => boxes[a]!.xmin + boxes[a]!.xmax - (boxes[b]!.xmin + boxes[b]!.xmax)
+      );
+      order.push(...line.items);
+    }
+  }
+  return order;
+}
+
+/**
+ * Reorders items carrying a `quad` into human reading order: multi-column inputs
+ * read column-by-column, single-column inputs line-by-line, and boxes within a
+ * line left-to-right. Detectors emit boxes in an arbitrary order, so detections
+ * and assembled blocks are ordered through this.
+ * @category Typescript API
+ * @param items The items to reorder, each carrying a `quad`.
+ * @returns The items in reading order.
+ */
+export function orderByReadingOrder<T extends { quad: readonly Point[] }>(items: T[]): T[] {
+  'worklet';
+  if (items.length <= 1) {
+    return items;
+  }
+  return readingOrder(items.map((it) => it.quad)).map((i) => items[i]!);
+}
+
+// A box wider than this multiple of its height is a horizontal line, never a
+// stacked-column glyph. A box joins a column when its x-span overlaps the
+// column's by COLUMN_X_OVERLAP of the narrower width and the y-gap is within
+// COLUMN_Y_GAP of its height.
+const COLUMN_GLYPH_ASPECT = 1.6;
+const COLUMN_X_OVERLAP = 0.25;
+const COLUMN_Y_GAP = 2.5;
+
+/**
+ * Clusters glyph-like, x-aligned, vertically-stacked boxes into columns; wide
+ * lines and isolated boxes are returned as `singles` to read normally. This lets
+ * a vertical-text pass add column reading without disturbing horizontal reads.
+ * @category Typescript API
+ * @param quads The detected text quads (ordered TL,TR,BR,BL).
+ * @returns The detected `columns` (each a top-to-bottom list of quads) and the
+ * leftover `singles` (horizontal lines / isolated boxes).
+ */
+export function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
+  columns: Point[][][];
+  singles: Point[][];
+} {
+  'worklet';
+  type Candidate = {
+    quad: Point[];
+    xmin: number;
+    xmax: number;
+    ymin: number;
+    ymax: number;
+    width: number;
+    height: number;
+  };
+  const candidates: Candidate[] = [];
+  const singles: Point[][] = [];
+  for (const q of quads) {
+    const { xmin, ymin, xmax, ymax } = boundsOfPoints(q);
+    const width = xmax - xmin;
+    const height = ymax - ymin;
+    if (width > height * COLUMN_GLYPH_ASPECT) {
+      singles.push(q as Point[]);
+    } else {
+      candidates.push({ quad: q as Point[], xmin, xmax, ymin, ymax, width, height });
+    }
+  }
+  // Grow each column top-to-bottom from its current bottom box, checking alignment
+  // against the column's accumulated x-range so a narrow glyph between wider ones
+  // doesn't break the run.
+  candidates.sort((a, b) => a.ymin - b.ymin);
+  type Column = { boxes: Candidate[]; xmin: number; xmax: number; bottom: number };
+  const columns: Column[] = [];
+  for (const box of candidates) {
+    let placed = false;
+    for (const column of columns) {
+      const overlap = Math.min(box.xmax, column.xmax) - Math.max(box.xmin, column.xmin);
+      const aligned = overlap > COLUMN_X_OVERLAP * Math.min(box.width, column.xmax - column.xmin);
+      const gap = box.ymin - column.bottom;
+      if (aligned && gap < COLUMN_Y_GAP * box.height && gap > -0.5 * box.height) {
+        column.boxes.push(box);
+        column.xmin = Math.min(column.xmin, box.xmin);
+        column.xmax = Math.max(column.xmax, box.xmax);
+        column.bottom = box.ymax;
+        placed = true;
+        break;
+      }
+    }
+    if (!placed) {
+      columns.push({ boxes: [box], xmin: box.xmin, xmax: box.xmax, bottom: box.ymax });
+    }
+  }
+  const grouped: Point[][][] = [];
+  for (const column of columns) {
+    if (column.boxes.length >= 2) {
+      grouped.push(column.boxes.map((b) => b.quad));
+    } else {
+      singles.push(column.boxes[0]!.quad);
+    }
+  }
+  return { columns: grouped, singles };
+}
+
+/**
+ * Collapses a greedy-CTC argmax path into recognized text and a confidence score:
+ * reserved low indices (the CTC blank, `< numSpecials`) and consecutive repeats
+ * are dropped from the text, and the confidence is the mean value over the
+ * non-reserved timesteps.
+ * @category Typescript API
+ * @param indices The per-timestep argmax indices.
+ * @param values The per-timestep max values (probabilities).
+ * @param charset The charset lookup, reserved tokens at the front.
+ * @param numSpecials Number of reserved low indices (default 1 = CTC blank).
+ * @returns The decoded `text` and its `confidence` in `[0, 1]`.
+ */
+export function ctcCollapse(
+  indices: number[],
+  values: number[],
+  charset: readonly string[],
+  numSpecials = 1
+): { text: string; confidence: number } {
+  'worklet';
+  let text = '';
+  let last = -1;
+  let sum = 0;
+  let count = 0;
+  for (let i = 0; i < indices.length; i++) {
+    const idx = indices[i]!;
+    if (idx >= numSpecials) {
+      sum += values[i]!;
+      count++;
+      if (idx !== last && idx < charset.length) {
+        text += charset[idx]!;
+      }
+    }
+    last = idx;
+  }
+  return { text, confidence: count === 0 ? 0 : sum / count };
+}
+
+/**
+ * Builds a CTC charset lookup: `numSpecials` reserved tokens (the CTC blank) are
+ * prepended, then the characters follow — a string is split into codepoints, an
+ * array is taken verbatim (preserving multi-codepoint entries) — so
+ * `charset[index]` decodes argmax `index`.
+ * @category Typescript API
+ * @param charset The model's ordered character set.
+ * @param numSpecials Number of reserved low indices (default 1 = CTC blank).
+ * @returns The charset lookup array, `numSpecials` reserved slots at the front.
+ */
+export function buildCharset(charset: string | readonly string[], numSpecials = 1): string[] {
+  'worklet';
+  const reserved = Array.from({ length: numSpecials }, (_unused, i) => `[reserved${i}]`);
+  const chars = typeof charset === 'string' ? Array.from(charset) : charset;
+  return [...reserved, ...chars];
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
new file mode 100644
index 0000000000..ce78efa018
--- /dev/null
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
@@ -0,0 +1,368 @@
+// OCR pipeline engine: the per-page / per-box worklet functions and the
+// construction-time contract resolvers behind `createOcr` (in ocr.ts). Internal
+// to the OCR task — not re-exported from the package index.
+//
+// Worklet source-order: the worklet plugin captures a referenced worklet by its
+// position in this file, so a worklet must be defined before any worklet that
+// calls it (recognizeCanvas -> recognizeQuad/recognizeGlyphStrip;
+// recognizeGlyphStrip -> readStackedColumn). The construction-time resolvers run
+// on the JS thread and have no such constraint.
+
+import { rnexecutorchJsi } from '../../../../native/bridge';
+import { tensor, type Tensor, type DType } from '../../../../core/tensor';
+import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
+import type { Model } from '../../../../core/model';
+
+import type { Point } from '../../ops/points';
+import {
+  resize,
+  cvtColor,
+  toChannelsFirst,
+  normalize,
+  warpQuad,
+  type ColorConversionCode,
+} from '../../ops/image';
+import { mapQuadToImage, orderQuad, quadSize, flattenQuad, splitTallQuad } from '../../ops/quad';
+import type { TextBoxExtractor } from './detectors';
+import { contentWidthFor, ctcCollapse, snapDetectBucket, snapRecognizeBucket } from './ocrUtils';
+
+// The detector consumes raw RGB scaled to [0,1]; its mean/std normalization is
+// baked into the model, so the client only divides by 255.
+const DETECTOR_ALPHA = 1 / 255;
+const DETECTOR_BETA = 0;
+
+// Per-timestep argmax + max value over recognizer logits `[..,T,V]`, computed
+// natively on the tensor buffer (avoids copying the whole tensor into JS). The
+// native op returns a flat [idx, value, ...] array, reshaped here.
+function ctcGreedyDecode(src: Tensor): { indices: number[]; values: number[] } {
+  'worklet';
+  const flat = rnexecutorchJsi.cv.ctcGreedyDecode(src) as number[];
+  const indices: number[] = [];
+  const values: number[] = [];
+  for (let i = 0; i < flat.length; i += 2) {
+    indices.push(flat[i]!);
+    values.push(flat[i + 1]!);
+  }
+  return { indices, values };
+}
+
+/** Per-detect-bucket scratch tensors, allocated once and reused across runs. */
+export type DetSet = {
+  readonly s: number;
+  readonly tColor: Tensor; // [s, s, 3]
+  readonly tCF: Tensor; // [3, s, s]
+  readonly tNorm: Tensor; // [3, s, s]
+  readonly tInput: Tensor; // [1, 3, s, s]
+  readonly tOutputs: readonly Tensor[];
+};
+
+/** The detector state, bundled so it can run on the full page or on a box crop. */
+export type DetectContext = {
+  readonly model: Model;
+  readonly detBuckets: readonly number[];
+  readonly numChannels: number;
+  readonly detCode: ColorConversionCode | null;
+  readonly extractBoxes: TextBoxExtractor;
+  readonly detSets: ReadonlyMap<number, DetSet>;
+};
+
+/** Per-recognize-width scratch tensors, allocated once and reused across runs. */
+export type RecSet = {
+  readonly width: number;
+  readonly tCanvas: Tensor;
+  readonly tCF: Tensor;
+  readonly tNorm: Tensor;
+  readonly tInput: Tensor;
+  readonly tLogits: Tensor;
+};
+
+/** The recognizer state; the source image is passed per call. */
+export type RecContext = {
+  readonly model: Model;
+  readonly recSetByWidth: ReadonlyMap<number, RecSet>;
+  readonly recBuckets: readonly number[];
+  readonly recH: number;
+  readonly charset: string[];
+  readonly normAlpha: number | readonly number[];
+  readonly normBeta: number | readonly number[];
+  readonly padValue: number;
+  // Optional custom decode; falls back to greedy CTC when absent.
+  readonly decode?: (
+    logits: Tensor,
+    charset: readonly string[]
+  ) => { readonly text: string; readonly confidence: number };
+};
+
+/** Extra state the vertical-text path needs: the detector and the source page. */
+export type VerticalContext = {
+  readonly detCtx: DetectContext;
+  readonly rawPage: Tensor;
+  readonly recC: number;
+  // Height/width ratio above which a box is treated as a stacked column.
+  readonly tallCropRatio: number;
+  // Per-page budget for the (expensive) stacked-column re-detection pass.
+  readonly redetectBudget: { remaining: number };
+};
+
+// Detects text boxes in `src` and returns their quads in `src` pixel space:
+// letterbox into the snapped square bucket, run the detector, and hand the raw
+// outputs to the model's extractor. `charLevel` requests per-glyph boxes for the
+// stacked-text pass.
+export function detectQuads(
+  ctx: DetectContext,
+  src: Tensor,
+  width: number,
+  height: number,
+  charLevel = false
+): Point[][] {
+  'worklet';
+  const detS = snapDetectBucket(width, height, ctx.detBuckets);
+  const detSet = ctx.detSets.get(detS)!;
+  // Only the source resize depends on the run's channel count; the rest is cached.
+  const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
+  try {
+    src
+      .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
+      .throughIf(ctx.detCode !== null, cvtColor, detSet.tColor, ctx.detCode!)
+      .through(toChannelsFirst, detSet.tCF)
+      .through(normalize, detSet.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
+      .copyTo(detSet.tInput);
+
+    ctx.model.execute(`detect_${detS}`, [detSet.tInput], [...detSet.tOutputs]);
+    const quads = ctx.extractBoxes(detSet.tOutputs, detS, charLevel);
+    return quads.map((q) => mapQuadToImage(q, detS, detS, width, height));
+  } finally {
+    tDetResize.dispose();
+  }
+}
+
+// Normalizes the already-warped recognizer canvas, runs the recognizer, and
+// decodes the logits — a custom decode if the model provides one, else greedy CTC
+// (the recognizer emits probabilities). Callers prepare `tCanvas` via `warpQuad`.
+function recognizeCanvas(
+  recCtx: RecContext,
+  recSet: RecSet,
+  bucketW: number
+): { text: string; conf: number } {
+  'worklet';
+  recSet.tCanvas
+    .through(toChannelsFirst, recSet.tCF)
+    .through(normalize, recSet.tNorm, { alpha: recCtx.normAlpha, beta: recCtx.normBeta })
+    .copyTo(recSet.tInput);
+  recCtx.model.execute(`recognize_${bucketW}`, [recSet.tInput], [recSet.tLogits]);
+  if (recCtx.decode) {
+    const r = recCtx.decode(recSet.tLogits, recCtx.charset);
+    return { text: r.text, conf: r.confidence };
+  }
+  const { indices, values } = ctcGreedyDecode(recSet.tLogits);
+  const { text, confidence } = ctcCollapse(indices, values, recCtx.charset);
+  return { text, conf: confidence };
+}
+
+// Recognizes one ordered (TL,TR,BR,BL) quad from `src`: snap its content width to
+// a recognizer bucket, warp it into the canvas, then recognize.
+export function recognizeQuad(
+  ctx: RecContext,
+  src: Tensor,
+  corners: readonly Point[]
+): { text: string; conf: number } {
+  'worklet';
+  const size = quadSize(corners);
+  const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
+  const desiredW = contentWidthFor(size.width, size.height, ctx.recH, maxRec);
+  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
+  const recSet = ctx.recSetByWidth.get(bucketW)!;
+  warpQuad(src, recSet.tCanvas, flattenQuad(corners), {
+    contentWidth: Math.min(desiredW, bucketW),
+    align: 'left',
+    padMode: 'constant',
+    padValue: ctx.padValue,
+  });
+  return recognizeCanvas(ctx, recSet, bucketW);
+}
+
+// Recognizes a sequence of glyph quads as a single line: each glyph is warped
+// upright to the recognizer height and placed side by side in one canvas (the
+// native warp composes directly, so there is no JS pixel assembly), then read in
+// one pass. A glyph box much taller than wide is first split into ~square
+// single-letter cells. Returns null when no usable text was produced.
+export function recognizeGlyphStrip(
+  recCtx: RecContext,
+  src: Tensor,
+  glyphs: readonly (readonly Point[])[]
+): { text: string; conf: number } | null {
+  'worklet';
+  const recH = recCtx.recH;
+  const maxRec = recCtx.recBuckets[recCtx.recBuckets.length - 1]!;
+  // Split any tall multi-letter box into single-letter cells and size each cell's
+  // warped width (aspect preserved) to lay out the strip.
+  const cells: { quad: readonly Point[]; width: number }[] = [];
+  let totalW = 0;
+  for (const glyph of glyphs) {
+    const glyphSize = quadSize(glyph);
+    if (glyphSize.width < 1 || glyphSize.height < 1) {
+      continue;
+    }
+    const parts = Math.max(1, Math.round(glyphSize.height / Math.max(1, glyphSize.width)));
+    for (const cell of splitTallQuad(glyph, parts)) {
+      const cellSize = quadSize(cell);
+      if (cellSize.width < 1 || cellSize.height < 1) {
+        continue;
+      }
+      const width = contentWidthFor(cellSize.width, cellSize.height, recH, maxRec);
+      cells.push({ quad: cell, width });
+      totalW += width;
+    }
+  }
+  if (cells.length === 0) {
+    return null;
+  }
+  const bucketW = snapRecognizeBucket(totalW, recCtx.recBuckets);
+  const recSet = recCtx.recSetByWidth.get(bucketW)!;
+  // Warp each cell into the canvas at its x-offset; the first warp clears + pads
+  // the whole canvas, the rest compose in with `clear: false`.
+  let xOff = 0;
+  for (let i = 0; i < cells.length; i++) {
+    if (xOff >= bucketW) {
+      break;
+    }
+    warpQuad(src, recSet.tCanvas, flattenQuad(cells[i]!.quad), {
+      contentWidth: cells[i]!.width,
+      offsetX: xOff,
+      clear: i === 0,
+      padMode: 'constant',
+      padValue: recCtx.padValue,
+    });
+    xOff += cells[i]!.width;
+  }
+  const { text, conf } = recognizeCanvas(recCtx, recSet, bucketW);
+  return text.length > 0 ? { text, conf } : null;
+}
+
+// Reads one tall box that the detector merged from several vertically-stacked
+// glyphs: crops it upright, re-detects the individual glyphs (char-level pass),
+// and recognizes them top-to-bottom via `recognizeGlyphStrip`. Returns null — the
+// caller then reads the box horizontally — when the box is too small, the per-page
+// re-detect budget is spent, or no glyphs are found.
+export function readStackedColumn(
+  recCtx: RecContext,
+  vctx: VerticalContext,
+  ordered: readonly Point[],
+  size: { width: number; height: number }
+): { text: string; conf: number } | null {
+  'worklet';
+  const boxW = Math.round(size.width);
+  const boxH = Math.round(size.height);
+  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
+    return null;
+  }
+  vctx.redetectBudget.remaining--;
+  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
+  // RGB conversion target — allocated lazily, only when the crop isn't RGB.
+  let tRecBox: Tensor | null = null;
+  try {
+    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
+      contentWidth: boxW,
+      align: 'left',
+      padMode: 'constant',
+      padValue: 0,
+    });
+    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
+    if (charQuads.length === 0) {
+      return null;
+    }
+    let boxSrc = tBoxRaw;
+    if (vctx.detCtx.detCode !== null) {
+      tRecBox = tensor('uint8', [boxH, boxW, vctx.recC]);
+      boxSrc = cvtColor(tBoxRaw, tRecBox, vctx.detCtx.detCode);
+    }
+    // Read the stack top-to-bottom by each glyph's upper edge.
+    const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
+    return recognizeGlyphStrip(recCtx, boxSrc, glyphs);
+  } finally {
+    tBoxRaw.dispose();
+    tRecBox?.dispose();
+  }
+}
+
+// Validates each `detect_<S>` method against the shared RGB `[1, 3, S, S]` input
+// contract and reads its (model-defined) output tensor specs from the model
+// metadata. Returns specs only — the task factory allocates and owns the tensors.
+// Runs at construction; throws on a contract mismatch.
+export function resolveDetectorContract(
+  model: Model,
+  detBuckets: readonly number[]
+): { s: number; outputs: { dtype: DType; shape: number[] }[] }[] {
+  return detBuckets.map((s) => {
+    const method = `detect_${s}`;
+    // Match the declared output count with wildcard specs so validateModelSchema
+    // enforces the RGB input contract without constraining the model's outputs.
+    const outCount = model.getMethodMeta(method).outputTensorMeta.length;
+    const meta = validateModelSchema(
+      model,
+      method,
+      [SymbolicTensor('float32', [1, 3, s, s])],
+      Array.from({ length: outCount }, () => SymbolicTensor())
+    );
+    return { s, outputs: meta.outputTensorMeta.map((m) => ({ dtype: m.dtype, shape: m.shape })) };
+  });
+}
+
+// Validates each `recognize_<W>` method and reads the recognizer contract: the
+// channel/height/vocab dimensions (constant across widths) plus each bucket's
+// input/output shapes. Returns specs only — the task factory allocates the
+// tensors. Runs at construction; throws on a contract mismatch.
+export function resolveRecognizerContract(
+  model: Model,
+  recBuckets: readonly number[]
+): {
+  recC: number;
+  recH: number;
+  vocabSize: number;
+  buckets: { width: number; inShape: number[]; outShape: number[] }[];
+} {
+  let recC = 0;
+  let recH = 0;
+  let vocabSize = 0;
+  const buckets = recBuckets.map((w, i) => {
+    const m = validateModelSchema(
+      model,
+      `recognize_${w}`,
+      [SymbolicTensor('float32', [1, 'C', 'H', 'W'])],
+      [SymbolicTensor('float32', [1, 'T', 'V'])]
+    );
+    const inShape = m.inputTensorMeta[0]!.shape;
+    const outShape = m.outputTensorMeta[0]!.shape;
+    if (i === 0) {
+      recC = inShape[1]!;
+      recH = inShape[2]!;
+      vocabSize = outShape[2]!;
+    }
+    return { width: w, inShape, outShape };
+  });
+  return { recC, recH, vocabSize, buckets };
+}
+
+// Frees a detector scratch-set's tensors.
+export function disposeDetSets(detSets: readonly DetSet[]): void {
+  for (const d of detSets) {
+    d.tColor.dispose();
+    d.tCF.dispose();
+    d.tNorm.dispose();
+    d.tInput.dispose();
+    for (const t of d.tOutputs) {
+      t.dispose();
+    }
+  }
+}
+
+// Frees a recognizer scratch-set's tensors.
+export function disposeRecSets(recSets: readonly RecSet[]): void {
+  for (const s of recSets) {
+    s.tCanvas.dispose();
+    s.tCF.dispose();
+    s.tNorm.dispose();
+    s.tInput.dispose();
+    s.tLogits.dispose();
+  }
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
deleted file mode 100644
index 330806d794..0000000000
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/supporting.ts
+++ /dev/null
@@ -1,302 +0,0 @@
-import type { WorkletRuntime } from 'react-native-worklets';
-
-import { tensor, type Tensor } from '../../../../core/tensor';
-import { loadModel } from '../../../../core/model';
-import { validateModelSchema, SymbolicTensor } from '../../../../core/modelSchema';
-import { wrapAsync } from '../../../../core/runtime';
-
-import type { ImageBuffer, ImageFormat } from '../../image';
-import { IMAGENET_NORM } from '../../../../constants';
-import { FORMAT_CHANNELS } from '../../ops/image';
-import { gridSample } from '../../ops/textBoxes';
-import { createImagePreprocessor } from '../preprocessing';
-import { argmaxRange } from './documentHelpers';
-
-// SLANet defaults; overridable per model via SupportingModel (the vocab's eos
-// token id and the structure decoder's max step count).
-const DEFAULT_EOS_TOKEN_ID = 49;
-const DEFAULT_MAX_STEPS = 501; // SLANet max_text_length (500) + 1
-
-// A dewarp grid estimated on a page without clear boundaries (e.g. text floating
-// on white, or a clean scan UVDoc wasn't trained on) can map most of the output
-// off the source, collapsing the page to near-blank and OCR to zero detections.
-// Guard: if the dewarped page keeps less than this fraction of the source's pixel
-// activity (variance), the warp is degenerate — keep the original page.
-const DEWARP_MIN_ACTIVITY_RATIO = 0.5;
-const DEWARP_ACTIVITY_STRIDE = 31; // subsample every Nth pixel for a cheap estimate
-
-// Variance of one channel sampled every DEWARP_ACTIVITY_STRIDE pixels — a cheap,
-// polarity-independent proxy for how much content (ink/edges) an image carries. A
-// blank/uniform page is ~0. Defined before createSupporting so the worklet plugin
-// captures it (a referenced worklet must precede its caller in source order).
-function sampledActivity(data: Uint8Array, channels: number): number {
-  'worklet';
-  let n = 0;
-  let sum = 0;
-  let sumSq = 0;
-  const step = channels * DEWARP_ACTIVITY_STRIDE;
-  for (let i = 0; i < data.length; i += step) {
-    const v = data[i]!;
-    sum += v;
-    sumSq += v * v;
-    n++;
-  }
-  if (n === 0) {
-    return 0;
-  }
-  const mean = sum / n;
-  return sumSq / n - mean * mean;
-}
-
-/**
- * Detected page orientation.
- * @category Types
- */
-export type Orientation = {
-  /** Degrees the page is rotated clockwise (0/90/180/270). Rotate by `-this` to correct. */
-  readonly rotationCW: 0 | 90 | 180 | 270;
-  readonly confidence: number;
-};
-
-/**
- * Recognized table structure.
- * @category Types
- */
-export type TableStructure = {
-  /** HTML `<tr>/<td>` skeleton (empty cells; fill by aligning OCR text). */
-  readonly html: string;
-  /** Raw SLANet structure token ids (sos/eos stripped). */
-  readonly tokens: number[];
-};
-
-/**
- * Model configuration for the fused PaddleOCR supporting models (orientation +
- * dewarp + table structure), all in one PTE.
- * @category Types
- */
-export type SupportingModel = {
-  readonly modelPath: string;
-  /** SLANet structure-token vocab (SLANET_STRUCTURE_VOCAB), index = token id. */
-  readonly vocab: readonly string[];
-  /** Token id that terminates AR decoding. Defaults to 49 (SLANet). */
-  readonly eosTokenId?: number;
-  /** Hard cap on AR decode steps. Defaults to 501 (SLANet max length + 1). */
-  readonly maxSteps?: number;
-};
-
-function tokensToHtml(tokens: number[], vocab: readonly string[], eosTokenId: number): string {
-  'worklet';
-  let html = '';
-  for (const t of tokens) {
-    if (t > 0 && t < eosTokenId && t < vocab.length) {
-      html += vocab[t]!;
-    }
-  }
-  return html;
-}
-
-/**
- * Creates the supporting-models runner (PaddleOCR fused helpers): page
- * orientation (PP-LCNet), geometric dewarp (UVDoc, applied in TS via the grid),
- * and table-structure recognition (SLANet_plus AR decode). One PTE, loaded once.
- * @category Typescript API
- * @param config Supporting-model path + table structure vocab.
- * @param runtime Optional worklet runtime thread.
- * @returns A promise resolving to the three capabilities + disposal controls.
- */
-export async function createSupporting(
-  config: SupportingModel,
-  runtime?: WorkletRuntime
-): Promise<{
-  dispose: () => void;
-  detectOrientation: (page: Tensor, format: ImageFormat) => Promise<Orientation>;
-  detectOrientationWorklet: (page: Tensor, format: ImageFormat) => Orientation;
-  dewarp: (page: Tensor, format: ImageFormat) => Promise<Tensor>;
-  dewarpWorklet: (page: Tensor, format: ImageFormat) => Tensor;
-  recognizeTable: (input: ImageBuffer) => Promise<TableStructure>;
-  recognizeTableWorklet: (input: ImageBuffer) => TableStructure;
-}> {
-  const { modelPath, vocab } = config;
-  const eosTokenId = config.eosTokenId ?? DEFAULT_EOS_TOKEN_ID;
-  const maxSteps = config.maxSteps ?? DEFAULT_MAX_STEPS;
-  const model = await wrapAsync(loadModel, runtime)(modelPath);
-
-  // --- orientation: image[1,3,224,224] (ImageNet) -> logits[1,4] ---
-  const oriMeta = validateModelSchema(
-    model,
-    'orientation',
-    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
-    [SymbolicTensor('float32', [1, 'K'])]
-  );
-  // --- dewarp: image[1,3,712,488] (/255) -> grid[1,2,gH,gW] ---
-  const dewMeta = validateModelSchema(
-    model,
-    'dewarp',
-    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
-    [SymbolicTensor('float32', [1, 2, 'gH', 'gW'])]
-  );
-  // --- table backbone + AR decoder ---
-  const encMeta = validateModelSchema(
-    model,
-    'table_encode',
-    [SymbolicTensor('float32', [1, 3, 'H', 'W'])],
-    [SymbolicTensor('float32', [1, 'C', 'F'])]
-  );
-  const decMeta = validateModelSchema(
-    model,
-    'table_decode_step',
-    [
-      SymbolicTensor('float32', [1, 'C', 'F']),
-      SymbolicTensor('float32', [1, 'H']),
-      SymbolicTensor('float32', [1, 'V']),
-    ],
-    [SymbolicTensor('float32', [1, 'V']), SymbolicTensor('float32', [1, 'H'])]
-  );
-
-  const oriShape = oriMeta.inputTensorMeta[0]!.shape;
-  const oriOutLen = oriMeta.outputTensorMeta[0]!.shape[1]!;
-  const dewShape = dewMeta.inputTensorMeta[0]!.shape;
-  const gridShape = dewMeta.outputTensorMeta[0]!.shape; // [1,2,gH,gW]
-  const tabShape = encMeta.inputTensorMeta[0]!.shape;
-  const featShape = encMeta.outputTensorMeta[0]!.shape;
-  const hidShape = decMeta.outputTensorMeta[1]!.shape;
-  const probShape = decMeta.outputTensorMeta[0]!.shape;
-  const hidLen = hidShape[1]!;
-  const vocabLen = probShape[1]!;
-
-  if (vocabLen !== vocab.length) {
-    throw new Error(
-      `Supporting: vocab length (${vocab.length}) must match the model's token dim (${vocabLen}).`
-    );
-  }
-
-  const orientationPreprocessor = createImagePreprocessor(
-    {
-      resizeMode: 'stretch',
-      interpolation: 'linear',
-      alpha: IMAGENET_NORM.alpha,
-      beta: IMAGENET_NORM.beta,
-    },
-    oriShape
-  );
-  const dewarpPreprocessor = createImagePreprocessor(
-    { resizeMode: 'stretch', interpolation: 'linear', alpha: 1 / 255, beta: 0 },
-    dewShape
-  );
-  const tablePreprocessor = createImagePreprocessor(
-    {
-      resizeMode: 'stretch',
-      interpolation: 'linear',
-      alpha: IMAGENET_NORM.alpha,
-      beta: IMAGENET_NORM.beta,
-    },
-    tabShape
-  );
-
-  const tensors = [
-    tensor('float32', oriMeta.outputTensorMeta[0]!.shape), // tOri
-    tensor('float32', gridShape), // tGrid
-    tensor('float32', featShape), // tFeatures
-    tensor('float32', hidShape), // tHidden
-    tensor('float32', probShape), // tOnehot
-    tensor('float32', probShape), // tProbs
-    tensor('float32', hidShape), // tNewHidden
-  ] as const;
-  const [tOri, tGrid, tFeatures, tHidden, tOnehot, tProbs, tNewHidden] = tensors;
-  const oriBuf = new Float32Array(oriOutLen);
-  const zeroHidden = new Float32Array(hidLen);
-  const zeroVocab = new Float32Array(vocabLen);
-  const onehotBuf = new Float32Array(vocabLen);
-  const probsBuf = new Float32Array(vocabLen);
-
-  const dispose = () => {
-    orientationPreprocessor.dispose();
-    dewarpPreprocessor.dispose();
-    tablePreprocessor.dispose();
-    tensors.forEach((t) => t.dispose());
-    model.dispose();
-  };
-
-  const detectOrientationWorklet = (page: Tensor, format: ImageFormat): Orientation => {
-    'worklet';
-    const tInput = orientationPreprocessor.processTensor(page, format);
-    model.execute('orientation', [tInput], [tOri]);
-    tOri.getData(oriBuf);
-    const cls = argmaxRange(oriBuf, 0, oriOutLen);
-    const best = oriBuf[cls]!;
-    let sumExp = 0;
-    for (let i = 0; i < oriOutLen; i++) {
-      sumExp += Math.exp(oriBuf[i]! - best);
-    }
-    const rotationCW = ((cls % 4) * 90) as 0 | 90 | 180 | 270;
-    const confidence = 1 / sumExp;
-    return { rotationCW, confidence };
-  };
-
-  // Dewarps the full-res page tensor in place: estimate the sampling field, apply
-  // it natively (cv::remap). Returns the dewarped tensor, or the input `page`
-  // unchanged when the warp is declined (caller owns whichever is returned).
-  const dewarpWorklet = (page: Tensor, format: ImageFormat): Tensor => {
-    'worklet';
-    const tInput = dewarpPreprocessor.processTensor(page, format);
-    model.execute('dewarp', [tInput], [tGrid]);
-    const h = page.shape[0]!;
-    const w = page.shape[1]!;
-    const ch = FORMAT_CHANNELS[format];
-    const tDst = tensor('uint8', [h, w, ch]);
-    try {
-      gridSample(page, tGrid, tDst);
-      const out = new Uint8Array(w * h * ch);
-      const src = new Uint8Array(w * h * ch);
-      tDst.getData(out);
-      page.getData(src);
-      // Degenerate-warp guard: a grid lacking page boundaries can push content
-      // off-canvas, leaving a near-blank page. If the dewarp collapsed the image's
-      // activity, decline it and keep the original (better an un-dewarped read than
-      // zero detections).
-      if (sampledActivity(out, ch) < DEWARP_MIN_ACTIVITY_RATIO * sampledActivity(src, ch)) {
-        tDst.dispose();
-        return page;
-      }
-      return tDst;
-    } catch (e) {
-      // On failure the caller can't see tDst to free it (success path returns it),
-      // so release it here before propagating.
-      tDst.dispose();
-      throw e;
-    }
-  };
-
-  const recognizeTableWorklet = (input: ImageBuffer): TableStructure => {
-    'worklet';
-    const tInput = tablePreprocessor.process(input);
-    model.execute('table_encode', [tInput], [tFeatures]);
-    tHidden.setData(zeroHidden);
-    tOnehot.setData(zeroVocab);
-    const tokens: number[] = [];
-    for (let step = 0; step < maxSteps; step++) {
-      model.execute('table_decode_step', [tFeatures, tHidden, tOnehot], [tProbs, tNewHidden]);
-      tProbs.getData(probsBuf);
-      const tok = argmaxRange(probsBuf, 0, vocabLen);
-      tokens.push(tok);
-      if (tok === eosTokenId) {
-        break;
-      }
-      tNewHidden.copyTo(tHidden);
-      onehotBuf.fill(0);
-      onehotBuf[tok] = 1;
-      tOnehot.setData(onehotBuf);
-    }
-    return { html: tokensToHtml(tokens, vocab, eosTokenId), tokens };
-  };
-
-  return {
-    dispose,
-    detectOrientation: wrapAsync(detectOrientationWorklet, runtime),
-    detectOrientationWorklet,
-    dewarp: wrapAsync(dewarpWorklet, runtime),
-    dewarpWorklet,
-    recognizeTable: wrapAsync(recognizeTableWorklet, runtime),
-    recognizeTableWorklet,
-  };
-}
diff --git a/packages/react-native-executorch/src/hooks/useDocumentOCR.ts b/packages/react-native-executorch/src/hooks/useDocumentOcr.ts
similarity index 58%
rename from packages/react-native-executorch/src/hooks/useDocumentOCR.ts
rename to packages/react-native-executorch/src/hooks/useDocumentOcr.ts
index a1e3bd87f5..beb1ecac91 100644
--- a/packages/react-native-executorch/src/hooks/useDocumentOCR.ts
+++ b/packages/react-native-executorch/src/hooks/useDocumentOcr.ts
@@ -1,6 +1,6 @@
 import { useModel } from './useModel';
 import { useResourceDownload } from './useResourceDownload';
-import { createDocumentOCR, type DocumentOCRModel } from '../extensions/cv/tasks/documentOCR';
+import { createDocumentOcr, type DocumentOcrModel } from '../extensions/cv/tasks/documentOcr';
 
 // Swap a model spec's hosted `modelPath` for its downloaded local path. Returns
 // undefined when the spec is absent (an optional model) or its path hasn't
@@ -12,53 +12,53 @@ const localize = <M extends { modelPath: string }>(
 
 /**
  * React hook for the document OCR pipeline: OCR + optional layout detection +
- * optional supporting (orientation/dewarp/table), assembled into reading-ordered
+ * optional document models (orientation/dewarp/table), assembled into reading-ordered
  * blocks. Downloads/compiles all enabled models, tracks progress and errors, and
  * cleans up native memory on unmount or config change.
  * @category Hooks
  * @typeParam L The type representing the layout region class labels.
- * @param config OCR model + optional layout/supporting models + flags.
+ * @param config OCR model + optional layout/document models + flags.
  * @param options Hook options.
  * @param options.preventLoad If true, prevents downloading and compiling the models.
  * @returns Loading state, error, download progress, and the document run functions.
  */
-export function useDocumentOCR<L>(
-  config: DocumentOCRModel<L>,
+export function useDocumentOcr<L>(
+  config: DocumentOcrModel<L>,
   options?: { preventLoad?: boolean }
 ) {
   const ocrDl = useResourceDownload(config.ocr.modelPath, options?.preventLoad);
   const layoutDl = useResourceDownload(config.layout?.modelPath, options?.preventLoad);
-  const supDl = useResourceDownload(config.supporting?.modelPath, options?.preventLoad);
+  const documentModelsDl = useResourceDownload(
+    config.documentModels?.modelPath,
+    options?.preventLoad
+  );
 
-  // Localize each enabled model; an optional model is "ready" when it's either
-  // absent or fully downloaded. Build the run config only once all are ready.
   const ocr = localize(config.ocr, ocrDl.localPath);
   const layout = localize(config.layout, layoutDl.localPath);
-  const supporting = localize(config.supporting, supDl.localPath);
-  const ready = !!ocr && (!config.layout || !!layout) && (!config.supporting || !!supporting);
-  const localConfig: DocumentOCRModel<L> | null = ready
-    ? { ...config, ocr: ocr!, layout, supporting }
+  const documentModels = localize(config.documentModels, documentModelsDl.localPath);
+  const ready =
+    !!ocr && (!config.layout || !!layout) && (!config.documentModels || !!documentModels);
+  const localConfig: DocumentOcrModel<L> | null = ready
+    ? { ...config, ocr: ocr!, layout, documentModels }
     : null;
 
-  const { model, error } = useModel(createDocumentOCR<L>, localConfig, [
+  const { model, error } = useModel(createDocumentOcr<L>, localConfig, [
     ocrDl.localPath,
     layoutDl.localPath,
-    supDl.localPath,
+    documentModelsDl.localPath,
   ]);
 
-  // Aggregate only the ENABLED downloads, so progress can't read 100% while a
-  // second/third model is still fetching.
   const downloads = [
     ocrDl,
     ...(config.layout ? [layoutDl] : []),
-    ...(config.supporting ? [supDl] : []),
+    ...(config.documentModels ? [documentModelsDl] : []),
   ];
 
   return {
     isReady: !!model,
     error: downloads.map((d) => d.downloadError).find(Boolean) || error,
     downloadProgress: Math.min(...downloads.map((d) => d.downloadProgress)),
-    runDocumentOCR: model?.runDocumentOCR,
-    runDocumentOCRWorklet: model?.runDocumentOCRWorklet,
+    runDocumentOcr: model?.runDocumentOcr,
+    runDocumentOcrWorklet: model?.runDocumentOcrWorklet,
   };
 }
diff --git a/packages/react-native-executorch/src/hooks/useOCR.ts b/packages/react-native-executorch/src/hooks/useOcr.ts
similarity index 75%
rename from packages/react-native-executorch/src/hooks/useOCR.ts
rename to packages/react-native-executorch/src/hooks/useOcr.ts
index 0699f4c1f1..e8b9dc3962 100644
--- a/packages/react-native-executorch/src/hooks/useOCR.ts
+++ b/packages/react-native-executorch/src/hooks/useOcr.ts
@@ -1,4 +1,4 @@
-import { createOCR, type OCRModel } from '../extensions/cv/tasks/ocr';
+import { createOcr, type OcrModel } from '../extensions/cv/tasks/ocr';
 import { useResourceDownload } from './useResourceDownload';
 import { useModel } from './useModel';
 
@@ -6,24 +6,24 @@ import { useModel } from './useModel';
  * React hook for running the unified OCR pipeline (EasyOCR / PaddleOCR).
  *
  * Downloads the fused PTE, instantiates the OCR task runner, and manages its
- * lifetime. Heavy work runs on a worklet thread; the returned `runOCR` resolves
+ * lifetime. Heavy work runs on a worklet thread; the returned `runOcr` resolves
  * with the recognized text regions.
  * @category Hooks
  * @param config OCR model configuration (one fused PTE + flat options). Use a
  * preset from `models.ocr.*`.
  * @param options Optional flags. `preventLoad` defers downloading/compiling the
  * model until set to `false`.
- * @returns Readiness flags, download progress, and the `runOCR` /
- * `runOCRWorklet` runners.
+ * @returns Readiness flags, download progress, and the `runOcr` /
+ * `runOcrWorklet` runners.
  */
-export function useOCR(config: OCRModel, options?: { preventLoad?: boolean }) {
+export function useOcr(config: OcrModel, options?: { preventLoad?: boolean }) {
   const { localPath, downloadProgress, downloadError } = useResourceDownload(
     config.modelPath,
     options?.preventLoad
   );
 
   const { model, error } = useModel(
-    createOCR,
+    createOcr,
     localPath ? { ...config, modelPath: localPath } : null,
     [localPath]
   );
@@ -33,7 +33,7 @@ export function useOCR(config: OCRModel, options?: { preventLoad?: boolean }) {
     error: downloadError || error,
     downloadProgress,
     localPath,
-    runOCR: model?.runOCR,
-    runOCRWorklet: model?.runOCRWorklet,
+    runOcr: model?.runOcr,
+    runOcrWorklet: model?.runOcrWorklet,
   };
 }
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index 8ea8514eeb..afcafa47f3 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -6,8 +6,8 @@ export * from './hooks/useInstanceSegmenter';
 export * from './hooks/useKeypointDetector';
 export * from './hooks/useObjectDetector';
 export * from './hooks/useTokenizer';
-export * from './hooks/useOCR';
-export * from './hooks/useDocumentOCR';
+export * from './hooks/useOcr';
+export * from './hooks/useDocumentOcr';
 export * from './hooks/useResourceDownload';
 export * from './hooks/useModel';
 
@@ -24,8 +24,9 @@ export * from './extensions/cv/tasks/keypointDetection';
 export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/nlp/tasks/tokenization';
 export * from './extensions/cv/tasks/ocr';
-export * from './extensions/cv/tasks/ocr/supporting';
-export * from './extensions/cv/tasks/documentOCR';
+export * from './extensions/cv/tasks/ocr/detectors';
+export type { DocumentModelsConfig } from './extensions/cv/tasks/ocr/documentModels';
+export * from './extensions/cv/tasks/documentOcr';
 
 // Core primitives — for library builders and power users
 export { tensor } from './core/tensor';
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index e9a06836ec..7a79f13329 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -4,8 +4,9 @@ import type { StyleTransferModel } from './extensions/cv/tasks/styleTransfer';
 import type { SemanticSegmentationModel } from './extensions/cv/tasks/semanticSegmentation';
 import type { KeypointDetectorModel } from './extensions/cv/tasks/keypointDetection';
 import type { InstanceSegmenterModel } from './extensions/cv/tasks/instanceSegmentation';
-import type { OCRModel, OCROptions } from './extensions/cv/tasks/ocr';
-import type { SupportingModel } from './extensions/cv/tasks/ocr/supporting';
+import type { OcrModel, OcrOptions } from './extensions/cv/tasks/ocr';
+import { craftExtractBoxes, dbnetExtractBoxes } from './extensions/cv/tasks/ocr/detectors';
+import type { DocumentModelsConfig } from './extensions/cv/tasks/ocr/documentModels';
 import {
   IMAGENET_NORM,
   IMAGENET1K_LABELS,
@@ -543,140 +544,141 @@ const ALL_MINILM_L6_V2_TOKENIZER = `${BASE_URL}-all-MiniLM-L6-v2/${VERSION_TAG}/
 // =============================================================================
 // OCR
 // =============================================================================
-const EASYOCR_OPTS: OCROptions = {
-  detectorKind: 'craft',
+const EASYOCR_OPTS: OcrOptions = {
+  extractBoxes: craftExtractBoxes,
   charset: alphabets.english,
   buckets: { detect: [800, 1280], recognize: [64, 128, 256, 512] },
 };
 
-const PADDLE_PPOCRV6_OPTS: OCROptions = {
-  detectorKind: 'dbnet',
+const PADDLE_PPOCRV6_OPTS: OcrOptions = {
+  extractBoxes: dbnetExtractBoxes,
+  dropScore: 0.5,
   charset: PPOCR_SYMBOLS,
   buckets: { detect: [640, 960, 1280], recognize: [160, 320, 480, 640, 1280] },
 };
 const OCR_REVISION = 'resolve/main';
 
 // English
-const EASYOCR_ENGLISH_XNNPACK: OCRModel = {
+const EASYOCR_ENGLISH_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
 };
-const EASYOCR_ENGLISH_COREML: OCRModel = {
+const EASYOCR_ENGLISH_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
 };
-const EASYOCR_ENGLISH_VULKAN: OCRModel = {
+const EASYOCR_ENGLISH_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/english/EasyOCR_english_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.english },
 };
 
 // Cyrillic
-const EASYOCR_CYRILLIC_XNNPACK: OCRModel = {
+const EASYOCR_CYRILLIC_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
 };
-const EASYOCR_CYRILLIC_COREML: OCRModel = {
+const EASYOCR_CYRILLIC_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
 };
-const EASYOCR_CYRILLIC_VULKAN: OCRModel = {
+const EASYOCR_CYRILLIC_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/cyrillic/EasyOCR_cyrillic_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.cyrillic },
 };
 
 // Latin
-const EASYOCR_LATIN_XNNPACK: OCRModel = {
+const EASYOCR_LATIN_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
 };
-const EASYOCR_LATIN_COREML: OCRModel = {
+const EASYOCR_LATIN_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
 };
-const EASYOCR_LATIN_VULKAN: OCRModel = {
+const EASYOCR_LATIN_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/latin/EasyOCR_latin_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.latin },
 };
 
 // Japanese
-const EASYOCR_JAPANESE_XNNPACK: OCRModel = {
+const EASYOCR_JAPANESE_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
 };
-const EASYOCR_JAPANESE_COREML: OCRModel = {
+const EASYOCR_JAPANESE_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
 };
-const EASYOCR_JAPANESE_VULKAN: OCRModel = {
+const EASYOCR_JAPANESE_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/japanese/EasyOCR_japanese_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.japanese },
 };
 
 // Simplified Chinese
-const EASYOCR_ZH_SIM_XNNPACK: OCRModel = {
+const EASYOCR_ZH_SIM_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
 };
-const EASYOCR_ZH_SIM_COREML: OCRModel = {
+const EASYOCR_ZH_SIM_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
 };
-const EASYOCR_ZH_SIM_VULKAN: OCRModel = {
+const EASYOCR_ZH_SIM_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/zh_sim/EasyOCR_zh_sim_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.zh_sim },
 };
 
 // Korean
-const EASYOCR_KOREAN_XNNPACK: OCRModel = {
+const EASYOCR_KOREAN_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
 };
-const EASYOCR_KOREAN_COREML: OCRModel = {
+const EASYOCR_KOREAN_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
 };
-const EASYOCR_KOREAN_VULKAN: OCRModel = {
+const EASYOCR_KOREAN_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/korean/EasyOCR_korean_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.korean },
 };
 
 // Telugu
-const EASYOCR_TELUGU_XNNPACK: OCRModel = {
+const EASYOCR_TELUGU_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
 };
-const EASYOCR_TELUGU_COREML: OCRModel = {
+const EASYOCR_TELUGU_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
 };
-const EASYOCR_TELUGU_VULKAN: OCRModel = {
+const EASYOCR_TELUGU_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/telugu/EasyOCR_telugu_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.telugu },
 };
 
 // Kannada
-const EASYOCR_KANNADA_XNNPACK: OCRModel = {
+const EASYOCR_KANNADA_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_xnnpack.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
 };
-const EASYOCR_KANNADA_COREML: OCRModel = {
+const EASYOCR_KANNADA_COREML: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_coreml.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
 };
-const EASYOCR_KANNADA_VULKAN: OCRModel = {
+const EASYOCR_KANNADA_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-easy-ocr/${OCR_REVISION}/kannada/EasyOCR_kannada_vulkan.pte`,
   ocrOpts: { ...EASYOCR_OPTS, charset: alphabets.kannada },
 };
 
-const PADDLE_PPOCRV6_XNNPACK: OCRModel = {
+const PADDLE_PPOCRV6_XNNPACK: OcrModel = {
   modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_xnnpack.pte`,
   ocrOpts: PADDLE_PPOCRV6_OPTS,
 };
-const PADDLE_PPOCRV6_COREML: OCRModel = {
+const PADDLE_PPOCRV6_COREML: OcrModel = {
   modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_coreml.pte`,
   ocrOpts: PADDLE_PPOCRV6_OPTS,
 };
-const PADDLE_PPOCRV6_VULKAN: OCRModel = {
+const PADDLE_PPOCRV6_VULKAN: OcrModel = {
   modelPath: `${BASE_URL}-pp-ocrv6/${OCR_REVISION}/PP-OCRv6_vulkan.pte`,
   ocrOpts: PADDLE_PPOCRV6_OPTS,
 };
@@ -708,23 +710,23 @@ const PP_DOCLAYOUT_VULKAN: ObjectDetectorModel<'xyxy', DocLayoutLabel> = {
 };
 
 // =============================================================================
-// Document helpers - Paddle supporting models
+// Document helper models - PaddleHelpers (orientation / dewarp / table structure)
 // =============================================================================
-const PP_TABLE_XNNPACK: SupportingModel = {
+const PP_HELPERS_XNNPACK: DocumentModelsConfig = {
   modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_xnnpack.pte`,
-  vocab: SLANET_STRUCTURE_VOCAB,
+  structureVocab: SLANET_STRUCTURE_VOCAB,
   eosTokenId: 49,
   maxSteps: 501,
 };
-const PP_TABLE_COREML: SupportingModel = {
+const PP_HELPERS_COREML: DocumentModelsConfig = {
   modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_coreml.pte`,
-  vocab: SLANET_STRUCTURE_VOCAB,
+  structureVocab: SLANET_STRUCTURE_VOCAB,
   eosTokenId: 49,
   maxSteps: 501,
 };
-const PP_TABLE_VULKAN: SupportingModel = {
+const PP_HELPERS_VULKAN: DocumentModelsConfig = {
   modelPath: `${BASE_URL}-paddle-helpers/${OCR_REVISION}/PaddleHelpers_vulkan.pte`,
-  vocab: SLANET_STRUCTURE_VOCAB,
+  structureVocab: SLANET_STRUCTURE_VOCAB,
   eosTokenId: 49,
   maxSteps: 501,
 };
@@ -991,11 +993,11 @@ export const models = {
       COREML: PP_DOCLAYOUT_COREML,
     },
   },
-  supporting: {
-    PP_SUPPORTING: {
-      XNNPACK: PP_TABLE_XNNPACK,
-      VULKAN: PP_TABLE_VULKAN,
-      COREML: PP_TABLE_COREML,
+  documentModels: {
+    PP_HELPERS: {
+      XNNPACK: PP_HELPERS_XNNPACK,
+      VULKAN: PP_HELPERS_VULKAN,
+      COREML: PP_HELPERS_COREML,
     },
   },
 };

From 8432dd7d5b46f6f2ee7012c5e0b1e8c3260499a1 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 00:16:43 +0200
Subject: [PATCH 26/29] [RNE Rewrite] chore: ocr screen cleanup, cspell words

---
 .cspell-wordlist.txt                          | 18 ++++++++++++++++
 apps/computer-vision/app/document/index.tsx   | 18 ++++++++--------
 apps/computer-vision/app/ocr/index.tsx        | 21 +++++++------------
 .../src/extensions/cv/tasks/ocr.ts            |  5 +----
 4 files changed, 35 insertions(+), 27 deletions(-)

diff --git a/.cspell-wordlist.txt b/.cspell-wordlist.txt
index aed3ee1e7d..43b009abf9 100644
--- a/.cspell-wordlist.txt
+++ b/.cspell-wordlist.txt
@@ -275,3 +275,21 @@ pcre
 libkleidicv
 thresholding
 binarization
+autoregressive
+dbnet
+dewarp
+dewarped
+dewarps
+doclayout
+easyocr
+letterboxed
+nums
+ocrv
+onehot
+ppocrv
+redetections
+svtr
+unclip
+unclips
+vctx
+lerp
diff --git a/apps/computer-vision/app/document/index.tsx b/apps/computer-vision/app/document/index.tsx
index 28b95fe50e..f2066c7db6 100644
--- a/apps/computer-vision/app/document/index.tsx
+++ b/apps/computer-vision/app/document/index.tsx
@@ -2,7 +2,7 @@ import React, { useState } from 'react';
 import { View, Text, StyleSheet, ScrollView, Switch, Platform } from 'react-native';
 import { commonStyles, ColorPalette } from '../../theme';
 import { useImage, Skia, ColorType, AlphaType, type SkImage } from '@shopify/react-native-skia';
-import { useDocumentOCR, models } from 'react-native-executorch';
+import { useDocumentOcr, models } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
 import { getImage } from '../../utils';
 import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
@@ -12,8 +12,8 @@ import { Button } from '../../components/Button';
 
 const PREVIEW_HEIGHT = 280;
 
-// Hosted per-backend model triplets (OCR + layout + supporting) — downloaded +
-// cached on-device from Hugging Face by `useDocumentOCR`. Backends are filtered
+// Hosted per-backend model triplets (OCR + layout + document models) — downloaded +
+// cached on-device from Hugging Face by `useDocumentOcr`. Backends are filtered
 // by platform (Vulkan = Android, CoreML = iOS, XNNPACK = both).
 type BackendKey = 'XNNPACK' | 'VULKAN' | 'COREML';
 const BACKENDS: { key: BackendKey; label: string; platforms: string[] }[] = [
@@ -100,16 +100,16 @@ function DocumentContent() {
 
   const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
 
-  // Hosted configs — `useDocumentOCR` downloads + caches each enabled model.
+  // Hosted configs — `useDocumentOcr` downloads + caches each enabled model.
   // orientation/dewarp are NOT baked here: they're passed per-run to
-  // `runDocumentOCR` below, so toggling them takes effect without a reload.
+  // `runDocumentOcr` below, so toggling them takes effect without a reload.
   const config = {
     ocr: models.ocr.PADDLE.PPOCRV6_SMALL[backend.key],
     ...(layoutOn ? { layout: models.layoutDetection.PP_DOCLAYOUT[backend.key] } : {}),
-    ...(supportingOn ? { supporting: models.supporting.PP_SUPPORTING[backend.key] } : {}),
+    ...(supportingOn ? { documentModels: models.documentModels.PP_HELPERS[backend.key] } : {}),
   };
 
-  const { isReady, downloadProgress, error: loadError, runDocumentOCR } = useDocumentOCR(config);
+  const { isReady, downloadProgress, error: loadError, runDocumentOcr } = useDocumentOcr(config);
 
   const handlePick = async (useCamera: boolean) => {
     setError(null);
@@ -127,14 +127,14 @@ function DocumentContent() {
   };
 
   const run = async () => {
-    if (!skiaImage || !runDocumentOCR) return;
+    if (!skiaImage || !runDocumentOcr) return;
     setIsProcessing(true);
     setError(null);
     try {
       const pixels = skiaImage.readPixels();
       if (!(pixels instanceof Uint8Array)) throw new Error('Expected Uint8Array from readPixels');
       const start = Date.now();
-      const out = await runDocumentOCR(
+      const out = await runDocumentOcr(
         {
           data: pixels,
           width: skiaImage.width(),
diff --git a/apps/computer-vision/app/ocr/index.tsx b/apps/computer-vision/app/ocr/index.tsx
index a0c3788521..1248ccfbca 100644
--- a/apps/computer-vision/app/ocr/index.tsx
+++ b/apps/computer-vision/app/ocr/index.tsx
@@ -2,7 +2,7 @@ import React, { useState } from 'react';
 import { View, Text, StyleSheet, ScrollView, Platform, Switch } from 'react-native';
 import { commonStyles, ColorPalette } from '../../theme';
 import { useImage } from '@shopify/react-native-skia';
-import { useOCR, models, type OCRDetection } from 'react-native-executorch';
+import { useOcr, models, type OcrDetection } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
 import { getImage } from '../../utils';
 import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
@@ -12,7 +12,7 @@ import { Button } from '../../components/Button';
 
 const PREVIEW_HEIGHT = 280;
 
-// Hosted PTEs — downloaded + cached on-device from Hugging Face by `useOCR`.
+// Hosted PTEs — downloaded + cached on-device from Hugging Face by `useOcr`.
 // Backends per platform: XNNPACK runs everywhere, Vulkan on Android, CoreML on iOS.
 const ALL_MODELS = [
   {
@@ -56,7 +56,7 @@ function OCRContent() {
   const [vertical, setVertical] = useState(false);
   const [imageUri, setImageUri] = useState<string | null>(null);
   const [isProcessing, setIsProcessing] = useState(false);
-  const [results, setResults] = useState<OCRDetection[]>([]);
+  const [results, setResults] = useState<OcrDetection[]>([]);
   const [wallMs, setWallMs] = useState<number | null>(null);
   const [error, setError] = useState<string | null>(null);
 
@@ -64,8 +64,8 @@ function OCRContent() {
 
   const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
 
-  // `useOCR` downloads + caches the hosted PTE from its Hugging Face URL.
-  const { isReady, downloadProgress, error: loadError, runOCR } = useOCR(selected.config);
+  // `useOcr` downloads + caches the hosted PTE from its Hugging Face URL.
+  const { isReady, downloadProgress, error: loadError, runOcr } = useOcr(selected.config);
 
   const handlePickImage = async (useCamera: boolean) => {
     setError(null);
@@ -82,7 +82,7 @@ function OCRContent() {
   };
 
   const runRecognition = async () => {
-    if (!skiaImage || !runOCR) return;
+    if (!skiaImage || !runOcr) return;
     setIsProcessing(true);
     setError(null);
     try {
@@ -99,7 +99,7 @@ function OCRContent() {
       };
       const start = Date.now();
       // `vertical` is a per-run option now — toggling it needs no model reload.
-      const output = await runOCR(buffer, { vertical });
+      const output = await runOcr(buffer, { vertical });
       setWallMs(Date.now() - start);
       setResults(output.detections);
     } catch (e: any) {
@@ -198,7 +198,6 @@ function OCRContent() {
                 {res.text}
               </Text>
               <View style={styles.resultMeta}>
-                <Text style={styles.resultMs}>{res.recognizeMs.toFixed(0)}ms</Text>
                 <Text style={styles.resultConfidence}>{Math.round(res.confidence * 100)}%</Text>
               </View>
             </View>
@@ -266,12 +265,6 @@ const styles = StyleSheet.create({
   tileUnit: { fontSize: 14, fontWeight: '600', color: '#6b73a3' },
   tileLabel: { fontSize: 11, color: '#868e96', marginTop: 4 },
   resultMeta: { flexDirection: 'row', alignItems: 'center' },
-  resultMs: {
-    fontSize: 12,
-    color: '#868e96',
-    marginRight: 10,
-    fontVariant: ['tabular-nums'],
-  },
   resultsContainer: {
     width: '100%',
     backgroundColor: '#fff',
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index f4d9d298ae..94bded8dc4 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -157,9 +157,6 @@ const TALL_CROP_RATIO = 1.5;
 const MAX_VERTICAL_REDETECTIONS = 8;
 // Vertical reads are lower-confidence and opt-in, so they skip the drop-score gate.
 const VERTICAL_DROP_SCORE = 0;
-// TEMP: stacked-column re-detection is disabled to measure whether char-level
-// column reading affects quality. Set true to restore it.
-const STACKED_COLUMNS_ENABLED = false;
 
 function pushDetection(
   out: OcrDetection[],
@@ -384,7 +381,7 @@ export async function createOcr(
       }
       for (const orderedQuad of singles) {
         const size = quadSize(orderedQuad);
-        if (STACKED_COLUMNS_ENABLED && size.height >= size.width * vctx.tallCropRatio) {
+        if (size.height >= size.width * vctx.tallCropRatio) {
           const stacked = readStackedColumn(recCtx, vctx, orderedQuad, size);
           if (stacked) {
             pushDetection(detections, VERTICAL_DROP_SCORE, stacked.text, stacked.conf, orderedQuad);

From 773eecf6948db2a244bee4e92fa43adccafe4505 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 08:44:42 +0200
Subject: [PATCH 27/29] =?UTF-8?q?[RNE=20Rewrite]=20refactor(ocr):=20audit?=
 =?UTF-8?q?=20cleanup=20=E2=80=94=20single=20bounds=20helper,=20naming,=20?=
 =?UTF-8?q?JSDoc?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .cspell-wordlist.txt                          |  1 -
 .../src/extensions/cv/ops/boxes.ts            | 27 +-------
 .../src/extensions/cv/tasks/documentOcr.ts    |  5 +-
 .../src/extensions/cv/tasks/ocr.ts            | 31 ++++-----
 .../cv/tasks/ocr/documentHelpers.ts           |  5 +-
 .../src/extensions/cv/tasks/ocr/pipeline.ts   | 65 +++++++------------
 6 files changed, 43 insertions(+), 91 deletions(-)

diff --git a/.cspell-wordlist.txt b/.cspell-wordlist.txt
index 43b009abf9..346d473ae4 100644
--- a/.cspell-wordlist.txt
+++ b/.cspell-wordlist.txt
@@ -291,5 +291,4 @@ redetections
 svtr
 unclip
 unclips
-vctx
 lerp
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index 361f69bc20..0d27b7a314 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,7 +1,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
-import { scalePoint, resizeFactors, type Point } from './points';
+import { scalePoint, resizeFactors } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
@@ -110,31 +110,6 @@ export function scaleBox<F extends BoxFormat>(
   }
 }
 
-/**
- * Computes the axis-aligned bounding box (`xyxy`) enclosing a set of points,
- * e.g. the corners of an oriented OCR {@link Quad}.
- * @category Utils
- * @param points The points to bound (need not be ordered).
- * @returns The enclosing bounding box in `xyxy` format.
- */
-export function boundingBoxOf(points: readonly Point[]): BoundingBox<'xyxy'> {
-  'worklet';
-  if (points.length === 0) {
-    return { format: 'xyxy', xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
-  }
-  let xmin = Infinity;
-  let ymin = Infinity;
-  let xmax = -Infinity;
-  let ymax = -Infinity;
-  for (const p of points) {
-    if (p.x < xmin) xmin = p.x;
-    if (p.y < ymin) ymin = p.y;
-    if (p.x > xmax) xmax = p.x;
-    if (p.y > ymax) ymax = p.y;
-  }
-  return { format: 'xyxy', xmin, ymin, xmax, ymax };
-}
-
 /**
  * Options for Non-Maximum Suppression (NMS).
  * @category Types
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
index d153162514..c565020d6c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
@@ -4,7 +4,8 @@ import { tensor } from '../../../core/tensor';
 import { wrapAsync } from '../../../core/runtime';
 import type { ImageBuffer } from '../image';
 import type { Point } from '../ops/points';
-import { boundingBoxOf, type BoundingBox } from '../ops/boxes';
+import type { BoundingBox } from '../ops/boxes';
+import { boundsOfPoints } from '../ops/quad';
 import { rotate, FORMAT_CHANNELS } from '../ops/image';
 import { createOcr, type OcrModel, type OcrDetection } from './ocr';
 import {
@@ -225,7 +226,7 @@ export async function createDocumentOcr<L>(
           ? [
               makeBlock<L>(
                 'ungrouped',
-                boundingBoxOf(detections.flatMap((d) => d.quad as Point[])),
+                { format: 'xyxy', ...boundsOfPoints(detections.flatMap((d) => d.quad as Point[])) },
                 1,
                 detections,
                 false
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
index 94bded8dc4..321dff054c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
@@ -22,8 +22,6 @@ import {
   readStackedColumn,
   resolveDetectorContract,
   resolveRecognizerContract,
-  disposeDetSets,
-  disposeRecSets,
   type DetSet,
   type RecSet,
   type DetectContext,
@@ -119,8 +117,8 @@ export type RunOcrOptions = {
 };
 
 /**
- * Model configuration required to instantiate an OCR task runner. One fused PTE
- * exposing `detect` + `recognize`.
+ * Model configuration required to instantiate an OCR task runner: one fused PTE
+ * exposing the per-size `detect_<S>` / `recognize_<W>` methods, plus its options.
  * @category Types
  */
 export type OcrModel = {
@@ -137,7 +135,7 @@ export type OcrDetection = {
   readonly confidence: number;
   /**
    * The oriented quad (TL,TR,BR,BL) in original image pixels. Derive the
-   * axis-aligned box with `boundingBoxOf(quad)` from `cv.ops.boxes` if needed.
+   * axis-aligned bounds with `boundsOfPoints(quad)` from `cv.ops.quad` if needed.
    */
   readonly quad: readonly Point[];
 };
@@ -172,13 +170,13 @@ function pushDetection(
 }
 
 /**
- * Creates a unified OCR runner for two-stage detect -> recognize models
- * (EasyOCR / PaddleOCR). It loads one fused PTE, validates the `detect` and
- * `recognize` methods, pre-allocates static scratch tensors sized from the
- * model's compiled shapes, and returns recognition + disposal controls.
+ * Creates a unified OCR runner for two-stage detect → recognize models
+ * (EasyOCR / PaddleOCR). It loads one fused PTE, validates every
+ * `detect_<S>` / `recognize_<W>` bucket method, pre-allocates static scratch
+ * tensors sized from the model's compiled shapes, and returns recognition +
+ * disposal controls.
  * @category Typescript API
- * @param config OCR task configuration containing the model path and flat
- * options.
+ * @param config The model path and OCR options ({@link OcrModel}).
  * @param runtime Optional worklet runtime thread on which to run the pipeline.
  * @returns A promise resolving to an object with recognition and disposal
  * controls.
@@ -276,8 +274,7 @@ export async function createOcr(
   }
 
   const dispose = () => {
-    disposeRecSets(recSets);
-    disposeDetSets(detSets);
+    allocated.forEach((t) => t.dispose());
     model.dispose();
   };
 
@@ -308,7 +305,7 @@ export async function createOcr(
       model,
       detBuckets,
       numChannels,
-      detCode: rgbCode,
+      toRgbCode: rgbCode,
       extractBoxes: ocrOpts.extractBoxes,
       detSets: detSetByS,
     };
@@ -339,7 +336,7 @@ export async function createOcr(
         padValue: recPadValue,
         decode: recDecode,
       };
-      const vctx: VerticalContext = {
+      const verticalCtx: VerticalContext = {
         detCtx,
         rawPage: tInputRaw,
         recC,
@@ -381,8 +378,8 @@ export async function createOcr(
       }
       for (const orderedQuad of singles) {
         const size = quadSize(orderedQuad);
-        if (size.height >= size.width * vctx.tallCropRatio) {
-          const stacked = readStackedColumn(recCtx, vctx, orderedQuad, size);
+        if (size.height >= size.width * verticalCtx.tallCropRatio) {
+          const stacked = readStackedColumn(recCtx, verticalCtx, orderedQuad, size);
           if (stacked) {
             pushDetection(detections, VERTICAL_DROP_SCORE, stacked.text, stacked.conf, orderedQuad);
             continue;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
index 39fe40c657..8f4e3f6d1a 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
@@ -1,6 +1,7 @@
 import type { ImageBuffer } from '../../image';
 import { FORMAT_CHANNELS } from '../../ops/image';
-import { boundingBoxOf, type BoundingBox } from '../../ops/boxes';
+import type { BoundingBox } from '../../ops/boxes';
+import { boundsOfPoints } from '../../ops/quad';
 import type { OcrDetection } from '../ocr';
 
 /**
@@ -105,7 +106,7 @@ export function fillTableCells(html: string, lines: readonly OcrDetection[]): st
   const centersX: number[] = [];
   const centersY: number[] = [];
   for (const line of lines) {
-    const box = boundingBoxOf(line.quad);
+    const box = boundsOfPoints(line.quad);
     centersX.push((box.xmin + box.xmax) / 2);
     centersY.push((box.ymin + box.ymax) / 2);
   }
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
index ce78efa018..316c1fd105 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
@@ -46,13 +46,16 @@ function ctcGreedyDecode(src: Tensor): { indices: number[]; values: number[] } {
   return { indices, values };
 }
 
-/** Per-detect-bucket scratch tensors, allocated once and reused across runs. */
+/**
+ * Per-detect-bucket scratch tensors, allocated once and reused across runs:
+ * `tColor [s,s,3]` → `tCF [3,s,s]` → `tNorm [3,s,s]` → `tInput [1,3,s,s]`.
+ */
 export type DetSet = {
   readonly s: number;
-  readonly tColor: Tensor; // [s, s, 3]
-  readonly tCF: Tensor; // [3, s, s]
-  readonly tNorm: Tensor; // [3, s, s]
-  readonly tInput: Tensor; // [1, 3, s, s]
+  readonly tColor: Tensor;
+  readonly tCF: Tensor;
+  readonly tNorm: Tensor;
+  readonly tInput: Tensor;
   readonly tOutputs: readonly Tensor[];
 };
 
@@ -61,7 +64,7 @@ export type DetectContext = {
   readonly model: Model;
   readonly detBuckets: readonly number[];
   readonly numChannels: number;
-  readonly detCode: ColorConversionCode | null;
+  readonly toRgbCode: ColorConversionCode | null;
   readonly extractBoxes: TextBoxExtractor;
   readonly detSets: ReadonlyMap<number, DetSet>;
 };
@@ -86,7 +89,7 @@ export type RecContext = {
   readonly normAlpha: number | readonly number[];
   readonly normBeta: number | readonly number[];
   readonly padValue: number;
-  // Optional custom decode; falls back to greedy CTC when absent.
+  /** Optional custom decode; falls back to greedy CTC when absent. */
   readonly decode?: (
     logits: Tensor,
     charset: readonly string[]
@@ -98,9 +101,9 @@ export type VerticalContext = {
   readonly detCtx: DetectContext;
   readonly rawPage: Tensor;
   readonly recC: number;
-  // Height/width ratio above which a box is treated as a stacked column.
+  /** Height/width ratio above which a box is treated as a stacked column. */
   readonly tallCropRatio: number;
-  // Per-page budget for the (expensive) stacked-column re-detection pass.
+  /** Per-page budget for the (expensive) stacked-column re-detection pass. */
   readonly redetectBudget: { remaining: number };
 };
 
@@ -123,7 +126,7 @@ export function detectQuads(
   try {
     src
       .through(resize, tDetResize, { mode: 'letterbox', interpolation: 'area', padValue: 0 })
-      .throughIf(ctx.detCode !== null, cvtColor, detSet.tColor, ctx.detCode!)
+      .throughIf(ctx.toRgbCode !== null, cvtColor, detSet.tColor, ctx.toRgbCode!)
       .through(toChannelsFirst, detSet.tCF)
       .through(normalize, detSet.tNorm, { alpha: DETECTOR_ALPHA, beta: DETECTOR_BETA })
       .copyTo(detSet.tInput);
@@ -246,35 +249,35 @@ export function recognizeGlyphStrip(
 // re-detect budget is spent, or no glyphs are found.
 export function readStackedColumn(
   recCtx: RecContext,
-  vctx: VerticalContext,
+  verticalCtx: VerticalContext,
   ordered: readonly Point[],
   size: { width: number; height: number }
 ): { text: string; conf: number } | null {
   'worklet';
   const boxW = Math.round(size.width);
   const boxH = Math.round(size.height);
-  if (boxW < 3 || boxH < 3 || vctx.redetectBudget.remaining <= 0) {
+  if (boxW < 3 || boxH < 3 || verticalCtx.redetectBudget.remaining <= 0) {
     return null;
   }
-  vctx.redetectBudget.remaining--;
-  const tBoxRaw = tensor('uint8', [boxH, boxW, vctx.detCtx.numChannels]);
+  verticalCtx.redetectBudget.remaining--;
+  const tBoxRaw = tensor('uint8', [boxH, boxW, verticalCtx.detCtx.numChannels]);
   // RGB conversion target — allocated lazily, only when the crop isn't RGB.
   let tRecBox: Tensor | null = null;
   try {
-    warpQuad(vctx.rawPage, tBoxRaw, flattenQuad(ordered), {
+    warpQuad(verticalCtx.rawPage, tBoxRaw, flattenQuad(ordered), {
       contentWidth: boxW,
       align: 'left',
       padMode: 'constant',
       padValue: 0,
     });
-    const charQuads = detectQuads(vctx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
+    const charQuads = detectQuads(verticalCtx.detCtx, tBoxRaw, boxW, boxH, /* charLevel */ true);
     if (charQuads.length === 0) {
       return null;
     }
     let boxSrc = tBoxRaw;
-    if (vctx.detCtx.detCode !== null) {
-      tRecBox = tensor('uint8', [boxH, boxW, vctx.recC]);
-      boxSrc = cvtColor(tBoxRaw, tRecBox, vctx.detCtx.detCode);
+    if (verticalCtx.detCtx.toRgbCode !== null) {
+      tRecBox = tensor('uint8', [boxH, boxW, verticalCtx.recC]);
+      boxSrc = cvtColor(tBoxRaw, tRecBox, verticalCtx.detCtx.toRgbCode);
     }
     // Read the stack top-to-bottom by each glyph's upper edge.
     const glyphs = charQuads.map((q) => orderQuad(q)).sort((a, b) => a[0]!.y - b[0]!.y);
@@ -342,27 +345,3 @@ export function resolveRecognizerContract(
   });
   return { recC, recH, vocabSize, buckets };
 }
-
-// Frees a detector scratch-set's tensors.
-export function disposeDetSets(detSets: readonly DetSet[]): void {
-  for (const d of detSets) {
-    d.tColor.dispose();
-    d.tCF.dispose();
-    d.tNorm.dispose();
-    d.tInput.dispose();
-    for (const t of d.tOutputs) {
-      t.dispose();
-    }
-  }
-}
-
-// Frees a recognizer scratch-set's tensors.
-export function disposeRecSets(recSets: readonly RecSet[]): void {
-  for (const s of recSets) {
-    s.tCanvas.dispose();
-    s.tCF.dispose();
-    s.tNorm.dispose();
-    s.tInput.dispose();
-    s.tLogits.dispose();
-  }
-}

From 507f7beca2bbed08fa83a313678b2ae7bd8c90c8 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 08:44:45 +0200
Subject: [PATCH 28/29] =?UTF-8?q?[RNE=20Rewrite]=20fix(cv):=20catch=20std:?=
 =?UTF-8?q?:exception=20=E2=80=94=20prebuilt=20OpenCV=20ships=20-fno-rtti?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../react-native-executorch/cpp/extensions/cv/image_ops.cpp   | 4 ++--
 .../react-native-executorch/cpp/extensions/cv/ocr_ops.cpp     | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
index eda54f2d15..69b08e0154 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.cpp
@@ -912,7 +912,7 @@ void install_warpByGrid(jsi::Runtime &rt, jsi::Object &module) {
         ::cv::Mat dstMat(h, w, cvType, dst->data_.get());
         try {
             ::cv::remap(srcMat, dstMat, mapX, mapY, ::cv::INTER_LINEAR, ::cv::BORDER_REPLICATE);
-        } catch (const ::cv::Exception &e) {
+        } catch (const std::exception &e) {
             throw jsi::JSError(rt, std::string("warpByGrid: OpenCV error: ") + e.what());
         }
         return jsi::Value(rt, args[2]);
@@ -1052,7 +1052,7 @@ void install_warpQuad(jsi::Runtime &rt, jsi::Object &module) {
                 content(::cv::Rect(0, 0, copyW, recH))
                     .copyTo(dstMat(::cv::Rect(offsetX, 0, copyW, recH)));
             }
-        } catch (const ::cv::Exception &e) {
+        } catch (const std::exception &e) {
             throw jsi::JSError(rt, std::string("warpQuad: OpenCV error: ") + e.what());
         }
         return jsi::Value(rt, args[1]);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index f11c2c2f83..eb8d936c39 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -536,7 +536,7 @@ void install_extractCraftTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                 static_cast<float>(opts.getProperty(rt, "linkThreshold").asNumber()),
                 static_cast<float>(opts.getProperty(rt, "lowTextThreshold").asNumber()),
                 restoreRatio, charLevel);
-        } catch (const ::cv::Exception &e) {
+        } catch (const std::exception &e) {
             throw jsi::JSError(rt, std::string("extractCraftTextBoxes: OpenCV error: ") + e.what());
         }
         return quadsToArray(rt, quads);
@@ -591,7 +591,7 @@ void install_extractDbnetTextBoxes(jsi::Runtime &rt, jsi::Object &module) {
                 static_cast<float>(opts.getProperty(rt, "unclipRatio").asNumber()),
                 static_cast<int32_t>(opts.getProperty(rt, "minBoxSide").asNumber()),
                 static_cast<int32_t>(opts.getProperty(rt, "maxCandidates").asNumber()));
-        } catch (const ::cv::Exception &e) {
+        } catch (const std::exception &e) {
             throw jsi::JSError(rt, std::string("extractDbnetTextBoxes: OpenCV error: ") + e.what());
         }
         return quadsToArray(rt, quads);

From 1177254dbe6ad2e7d78e66211258dc3bdd9f6615 Mon Sep 17 00:00:00 2001
From: benITo47 <bartlomiej.obrochta03@gmail.com>
Date: Thu, 2 Jul 2026 12:50:44 +0200
Subject: [PATCH 29/29] =?UTF-8?q?[RNE=20Rewrite]=20refactor(ocr):=20review?=
 =?UTF-8?q?=20cleanup=20=E2=80=94=20ocr=20family=20folder,=20slimmer=20nat?=
 =?UTF-8?q?ive=20contract,=20generic=20bounds?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .cspell-wordlist.txt                          |   1 -
 .../cpp/extensions/cv/box_ops.cpp             |  36 ++---
 .../cpp/extensions/cv/image_ops.h             |   3 -
 .../cpp/extensions/cv/ocr_ops.cpp             |  50 +------
 .../cpp/extensions/cv/ocr_ops.h               |   9 --
 .../react-native-executorch/src/core/model.ts |   7 +-
 .../src/extensions/cv/ops/boxes.ts            |  16 ++-
 .../src/extensions/cv/ops/points.ts           |  47 ++-----
 .../src/extensions/cv/ops/quad.ts             |  64 +++++----
 .../src/extensions/cv/tasks/ocr/detectors.ts  |  53 +++----
 .../cv/tasks/ocr/documentHelpers.ts           | 126 -----------------
 .../extensions/cv/tasks/ocr/documentModels.ts | 133 +++++++++++++++++-
 .../cv/tasks/{ => ocr}/documentOcr.ts         |  31 ++--
 .../src/extensions/cv/tasks/{ => ocr}/ocr.ts  |  41 +++---
 .../src/extensions/cv/tasks/ocr/ocrUtils.ts   | 119 +++++-----------
 .../src/extensions/cv/tasks/ocr/pipeline.ts   |   8 +-
 .../src/hooks/useDocumentOcr.ts               |   2 +-
 .../src/hooks/useOcr.ts                       |   2 +-
 packages/react-native-executorch/src/index.ts |   4 +-
 .../react-native-executorch/src/models.ts     |   2 +-
 20 files changed, 322 insertions(+), 432 deletions(-)
 delete mode 100644 packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/documentOcr.ts (93%)
 rename packages/react-native-executorch/src/extensions/cv/tasks/{ => ocr}/ocr.ts (92%)

diff --git a/.cspell-wordlist.txt b/.cspell-wordlist.txt
index 346d473ae4..3d35595f05 100644
--- a/.cspell-wordlist.txt
+++ b/.cspell-wordlist.txt
@@ -291,4 +291,3 @@ redetections
 svtr
 unclip
 unclips
-lerp
diff --git a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
index 031a5c948b..59315641a2 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/box_ops.cpp
@@ -166,28 +166,20 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
         std::vector<std::vector<std::int32_t>> groups;
         std::vector<bool> suppressed(candidates.size(), false);
 
-        // Decode every candidate's box to xyxy + area once, indexed by candidate
-        // position, so the O(N^2) suppression loop below reads them instead of
-        // re-decoding box j on every pass.
-        struct DecodedBox {
-            float xmin, ymin, xmax, ymax, area;
-        };
-        std::vector<DecodedBox> decoded(candidates.size());
+        // Decode every candidate to xyxy once, not per pair in the O(n²) loop.
+        std::vector<std::array<float, 4>> decoded(candidates.size());
         for (size_t k = 0; k < candidates.size(); ++k) {
             const std::int32_t idx = candidates[k].first;
-            auto [xmin, ymin, xmax, ymax] = decodeToXyxy(
+            decoded[k] = decodeToXyxy(
                 boxesPtr[idx * 4 + 0],
                 boxesPtr[idx * 4 + 1],
                 boxesPtr[idx * 4 + 2],
                 boxesPtr[idx * 4 + 3],
                 boxFormat);
-            decoded[k] = {
-                .xmin = xmin,
-                .ymin = ymin,
-                .xmax = xmax,
-                .ymax = ymax,
-                .area = (xmax - xmin) * (ymax - ymin)};
         }
+        const auto boxArea = [](const std::array<float, 4> &box) {
+            return (box[2] - box[0]) * (box[3] - box[1]);
+        };
 
         for (size_t i = 0; i < candidates.size(); ++i) {
             if (suppressed[i]) {
@@ -195,8 +187,8 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
             }
 
             const std::int32_t idxI = candidates[i].first;
-            const DecodedBox &a = decoded[i];
-            const float areaA = a.area;
+            const auto &[aXmin, aYmin, aXmax, aYmax] = decoded[i];
+            const float areaA = boxArea(decoded[i]);
 
             std::vector<std::int32_t> overlapping = {idxI};
 
@@ -206,13 +198,13 @@ void install_nms(jsi::Runtime &rt, jsi::Object &module) {
                 }
 
                 const std::int32_t idxJ = candidates[j].first;
-                const DecodedBox &b = decoded[j];
-                const float areaB = b.area;
+                const auto &[bXmin, bYmin, bXmax, bYmax] = decoded[j];
+                const float areaB = boxArea(decoded[j]);
 
-                const float interYMin = std::max(a.ymin, b.ymin);
-                const float interXMin = std::max(a.xmin, b.xmin);
-                const float interYMax = std::min(a.ymax, b.ymax);
-                const float interXMax = std::min(a.xmax, b.xmax);
+                const float interYMin = std::max(aYmin, bYmin);
+                const float interXMin = std::max(aXmin, bXmin);
+                const float interYMax = std::min(aYmax, bYmax);
+                const float interXMax = std::min(aXmax, bXmax);
 
                 const float interH = std::max(0.0f, interYMax - interYMin);
                 const float interW = std::max(0.0f, interXMax - interXMin);
diff --git a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
index a694809aac..6a4235f3bc 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/image_ops.h
@@ -10,9 +10,6 @@ void install_toChannelsLast(facebook::jsi::Runtime &rt, facebook::jsi::Object &m
 void install_normalize(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_applyColormap(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 void install_rotate(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-// Warp `src` through a low-res [-1,1] sampling grid into `dst` (dewarp / remap).
 void install_warpByGrid(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-// Perspective-crop an oriented quad of `src` into the `dst` canvas (crop +
-// resize-to-height + pad/align).
 void install_warpQuad(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::image_ops
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
index eb8d936c39..1868ca0365 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.cpp
@@ -21,10 +21,6 @@
 #include "core/dtype.h"
 #include "core/tensor.h"
 
-// Detector postprocessing geometry: CRAFT text-map grouping + DBNet prob-map ->
-// oriented quads. Pure OpenCV, kept native. ctcGreedyDecode (per-timestep argmax
-// + max prob) is native too; the CTC blank-collapse, charset mapping, and
-// confidence aggregation stay in TypeScript.
 namespace rnexecutorch::extensions::cv::ocr_ops {
 namespace jsi = facebook::jsi;
 using TensorHostObject = rnexecutorch::core::tensor::TensorHostObject;
@@ -41,7 +37,6 @@ struct Box {
 struct Quad {
     std::array<::cv::Point2f, 4> pts;
     float score = 1.0f;
-    float angle = 0.0f;
 };
 
 float dist(const ::cv::Point2f &a, const ::cv::Point2f &b) {
@@ -307,32 +302,14 @@ std::vector<Box> groupTextBoxes(std::vector<Box> boxes, float centerThreshold,
         merged.push_back(current);
     }
 
-    // remove small boxes
+    // Remove small boxes. Output order is unspecified — the TypeScript pipeline
+    // derives reading order geometrically for every result set.
     std::vector<Box> filtered;
     for (const auto &b : merged) {
         if (minSide(b) > minSideThreshold && maxSide(b) > maxSideThreshold) {
             filtered.push_back(b);
         }
     }
-
-    // reading order: rows by top-Y, then left-to-right within a row
-    std::ranges::sort(filtered,
-                      [](const Box &a, const Box &b) { return a.y0 < b.y0; });
-    float yThresh = 0.0f;
-    if (!filtered.empty()) {
-        float total = 0.0f;
-        for (const auto &b : filtered) {
-            total += minSide(b);
-        }
-        yThresh = (total / static_cast<float>(filtered.size())) * 0.5f;
-    }
-    for (auto rowBegin = filtered.begin(); rowBegin != filtered.end();) {
-        const float rowY = rowBegin->y0;
-        auto rowEnd = std::find_if(rowBegin, filtered.end(),
-                                   [rowY, yThresh](const Box &b) { return b.y0 - rowY > yThresh; });
-        std::sort(rowBegin, rowEnd, [](const Box &a, const Box &b) { return a.x0 < b.x0; });
-        rowBegin = rowEnd;
-    }
     return filtered;
 }
 
@@ -368,7 +345,6 @@ std::vector<Quad> extractCraft(float *data, int32_t heatW, int32_t heatH, float
     for (const auto &b : boxes) {
         Quad q;
         q.score = 1.0f;
-        q.angle = b.angle;
         // De-skew near-horizontal lines by rotating the AABB corners about the
         // center. A near-vertical line (angle ~ -90, from a tall/stacked region)
         // is NOT flipped flat — that would lay an upright column on its side and
@@ -435,7 +411,6 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &prob, float binThreshold, float
         expanded.points(c.data());
         Quad q;
         q.score = score;
-        q.angle = expanded.angle;
         auto minX = static_cast<float>(w);
         auto minY = static_cast<float>(h);
         float maxX = 0;
@@ -454,26 +429,14 @@ std::vector<Quad> extractDbnet(const ::cv::Mat &prob, float binThreshold, float
         }
         quads.push_back(q);
     }
-
-    // Reading order: top -> bottom by ~row, then left -> right. Quantise y into
-    // row bands first so the comparator is a valid strict-weak ordering — a raw
-    // `|dy| > threshold` test is intransitive (a~b, b~c, but a<c) and aborts under
-    // libc++ hardening.
-    constexpr float kRowBand = 10.0f;
-    std::ranges::sort(quads, [](const Quad &a, const Quad &b) {
-        const int rowA = static_cast<int>(std::floor(a.pts[0].y / kRowBand));
-        const int rowB = static_cast<int>(std::floor(b.pts[0].y / kRowBand));
-        if (rowA != rowB) {
-            return rowA < rowB;
-        }
-        return a.pts[0].x < b.pts[0].x;
-    });
+    // Output order is unspecified — the TypeScript pipeline derives reading
+    // order geometrically for every result set.
     return quads;
 }
 
-// Flatten quads to a JS double array, 10 per box (x0,y0..x3,y3,score,angle).
+// Flatten quads to a JS double array, 9 per box (x0,y0..x3,y3,score).
 jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
-    jsi::Array out(rt, quads.size() * 10);
+    jsi::Array out(rt, quads.size() * 9);
     size_t idx = 0;
     for (const auto &q : quads) {
         for (std::size_t k = 0; k < 4; ++k) {
@@ -481,7 +444,6 @@ jsi::Array quadsToArray(jsi::Runtime &rt, const std::vector<Quad> &quads) {
             out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.pts[k].y)));
         }
         out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.score)));
-        out.setValueAtIndex(rt, idx++, jsi::Value(static_cast<double>(q.angle)));
     }
     return out;
 }
diff --git a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
index 002fb64601..1e47193bb0 100644
--- a/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
+++ b/packages/react-native-executorch/cpp/extensions/cv/ocr_ops.h
@@ -3,16 +3,7 @@
 #include <jsi/jsi.h>
 
 namespace rnexecutorch::extensions::cv::ocr_ops {
-// CRAFT half-res region+affinity heatmap [..,Hd,Wd,2] -> flat array of oriented
-// quads (10 doubles/box: x0,y0..x3,y3,score,angle), in detector-input pixels.
 void install_extractCraftTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-
-// DBNet full-res probability map [..,H,W] -> flat array of oriented quads (10
-// doubles/box), in detector-input pixels.
 void install_extractDbnetTextBoxes(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
-
-// Per-timestep argmax + max value over [..,T,V] logits -> flat [idx,value,...].
-// `value` is the raw max activation; softmax the tensor beforehand if a caller
-// needs a probability.
 void install_ctcGreedyDecode(facebook::jsi::Runtime &rt, facebook::jsi::Object &module);
 } // namespace rnexecutorch::extensions::cv::ocr_ops
diff --git a/packages/react-native-executorch/src/core/model.ts b/packages/react-native-executorch/src/core/model.ts
index 5b7274d6ec..3b5b133476 100644
--- a/packages/react-native-executorch/src/core/model.ts
+++ b/packages/react-native-executorch/src/core/model.ts
@@ -122,12 +122,7 @@ export interface Model {
   /**
    * Unloads a single previously-executed method, freeing its memory-planned
    * activation arena (and, on graph-compiling backends like CoreML, its
-   * compiled graph). The method transparently reloads on its next `execute`.
-   *
-   * Use this to bound native memory when many distinct methods are executed
-   * over a session — e.g. bucketed OCR, where each `detect_<S>`/`recognize_<W>`
-   * size that is ever run would otherwise stay resident for the model's
-   * lifetime.
+   * compiled graph). The method reloads on its next `execute`.
    * @param methodName The exported method to unload.
    * @returns `true` if a loaded method was freed, `false` if it was not loaded
    * (a harmless no-op).
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
index 0d27b7a314..0ce289be00 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/boxes.ts
@@ -1,7 +1,7 @@
 import { rnexecutorchJsi } from '../../../native/bridge';
 import type { Tensor } from '../../../core/tensor';
 import type { ResizeMode } from './image';
-import { scalePoint, resizeFactors } from './points';
+import { scalePoint } from './points';
 
 /**
  * Mapping of bounding box formats to their coordinate representations.
@@ -73,8 +73,8 @@ export function scaleBox<F extends BoxFormat>(
   }
 ): BoundingBox<F> {
   'worklet';
-  const { scaleX, scaleY } = resizeFactors(opts.from, opts.to, opts.resizeMode);
-
+  // Both resize maps are affine, so a span scales exactly as the difference of
+  // two mapped points.
   switch (box.format) {
     case 'xyxy': {
       const pMin = scalePoint({ x: box.xmin, y: box.ymin }, opts);
@@ -89,22 +89,24 @@ export function scaleBox<F extends BoxFormat>(
     }
     case 'xywh': {
       const pMin = scalePoint({ x: box.xmin, y: box.ymin }, opts);
+      const pFar = scalePoint({ x: box.xmin + box.w, y: box.ymin + box.h }, opts);
       return {
         format: 'xywh',
         xmin: pMin.x,
         ymin: pMin.y,
-        w: box.w / scaleX,
-        h: box.h / scaleY,
+        w: pFar.x - pMin.x,
+        h: pFar.y - pMin.y,
       } as BoundingBox<F>;
     }
     case 'cxcywh': {
       const pCenter = scalePoint({ x: box.cx, y: box.cy }, opts);
+      const pFar = scalePoint({ x: box.cx + box.w, y: box.cy + box.h }, opts);
       return {
         format: 'cxcywh',
         cx: pCenter.x,
         cy: pCenter.y,
-        w: box.w / scaleX,
-        h: box.h / scaleY,
+        w: pFar.x - pCenter.x,
+        h: pFar.y - pCenter.y,
       } as BoundingBox<F>;
     }
   }
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/points.ts b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
index 47cbeedfa4..fa803dd4a2 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/points.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/points.ts
@@ -10,37 +10,9 @@ export type Point = {
 };
 
 /**
- * Per-axis scale and offset that map a `to`-space coordinate back into
- * `from`-space, for an aspect-preserving letterbox or an axis stretch. The
- * inverse map is `(coord − offset) / scale` per axis (offset is 0 for stretch).
- * Shared by {@link scalePoint} and `scaleBox` so the factors are derived once.
- * @category Utils
- * @param from The source bounds (e.g. model input dimensions).
- * @param to The destination bounds (e.g. original image dimensions).
- * @param resizeMode The resize mode the source was produced with.
- * @returns The per-axis `scaleX`/`scaleY` and `offsetX`/`offsetY`.
- */
-export function resizeFactors(
-  from: { readonly width: number; readonly height: number },
-  to: { readonly width: number; readonly height: number },
-  resizeMode: Exclude<ResizeMode, 'crop'>
-): { scaleX: number; scaleY: number; offsetX: number; offsetY: number } {
-  'worklet';
-  if (resizeMode === 'letterbox') {
-    const scale = Math.min(from.width / to.width, from.height / to.height);
-    return {
-      scaleX: scale,
-      scaleY: scale,
-      offsetX: (from.width - to.width * scale) / 2.0,
-      offsetY: (from.height - to.height * scale) / 2.0,
-    };
-  }
-  return { scaleX: from.width / to.width, scaleY: from.height / to.height, offsetX: 0, offsetY: 0 };
-}
-
-/**
- * Helper function to scale a 2D point based on resize mode and resolution
- * changes.
+ * Maps a `from`-space coordinate (e.g. model input pixels) back into `to`-space
+ * (e.g. original image pixels), inverting the aspect-preserving letterbox or the
+ * per-axis stretch the source was produced with.
  * @category Utils
  * @param point The original coordinate point to scale.
  * @param opts Options detailing the scaling factors and resize mode.
@@ -59,6 +31,15 @@ export function scalePoint(
   }
 ): Point {
   'worklet';
-  const { scaleX, scaleY, offsetX, offsetY } = resizeFactors(opts.from, opts.to, opts.resizeMode);
-  return { x: (point.x - offsetX) / scaleX, y: (point.y - offsetY) / scaleY };
+  const { from, to } = opts;
+  switch (opts.resizeMode) {
+    case 'letterbox': {
+      const scale = Math.min(from.width / to.width, from.height / to.height);
+      const offsetX = (from.width - to.width * scale) / 2;
+      const offsetY = (from.height - to.height * scale) / 2;
+      return { x: (point.x - offsetX) / scale, y: (point.y - offsetY) / scale };
+    }
+    case 'stretch':
+      return { x: (point.x * to.width) / from.width, y: (point.y * to.height) / from.height };
+  }
 }
diff --git a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
index 99ebb01daa..ad2071d051 100644
--- a/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
+++ b/packages/react-native-executorch/src/extensions/cv/ops/quad.ts
@@ -1,45 +1,41 @@
 import { scalePoint, type Point } from './points';
+import type { BoundingBox, BoxFormat } from './boxes';
 
 /**
  * An oriented quadrilateral in pixel space: `points` are the four corners ordered
- * top-left, top-right, bottom-right, bottom-left, `score` is the region confidence
- * in `[0, 1]`, and `angle` is the rotation in degrees.
+ * top-left, top-right, bottom-right, bottom-left, and `score` is the region
+ * confidence in `[0, 1]`. Orientation lives in the corners themselves.
  * @category Types
  */
 export type Quad = {
   readonly points: readonly Point[];
   readonly score: number;
-  readonly angle: number;
 };
 
-/**
- * The axis-aligned bounds of a set of points.
- * @category Types
- */
-export type Bounds = { xmin: number; ymin: number; xmax: number; ymax: number };
-
 const distance = (a: Point, b: Point): number => {
   'worklet';
   return Math.hypot(b.x - a.x, b.y - a.y);
 };
 
-const lerp = (a: Point, b: Point, t: number): Point => {
+const interpolatePoint = (a: Point, b: Point, t: number): Point => {
   'worklet';
   return { x: a.x + (b.x - a.x) * t, y: a.y + (b.y - a.y) * t };
 };
 
 /**
- * Computes the axis-aligned bounds enclosing a set of points. Returns a zero box
- * for empty input.
+ * Computes the axis-aligned bounding box enclosing a set of points, in the
+ * requested box format. Returns a zero box for empty input.
  * @category Typescript API
+ * @typeParam F Bounding box coordinate format.
  * @param points The points to enclose.
- * @returns The enclosing `{ xmin, ymin, xmax, ymax }` bounds.
+ * @param format The coordinate format of the returned box.
+ * @returns The enclosing {@link BoundingBox} in `format`.
  */
-export function boundsOfPoints(points: readonly Point[]): Bounds {
+export function boundsOfPoints<F extends BoxFormat>(
+  points: readonly Point[],
+  format: F
+): BoundingBox<F> {
   'worklet';
-  if (points.length === 0) {
-    return { xmin: 0, ymin: 0, xmax: 0, ymax: 0 };
-  }
   let xmin = Infinity;
   let ymin = Infinity;
   let xmax = -Infinity;
@@ -50,7 +46,23 @@ export function boundsOfPoints(points: readonly Point[]): Bounds {
     if (p.x > xmax) xmax = p.x;
     if (p.y > ymax) ymax = p.y;
   }
-  return { xmin, ymin, xmax, ymax };
+  if (points.length === 0) {
+    xmin = ymin = xmax = ymax = 0;
+  }
+  switch (format) {
+    case 'xyxy':
+      return { format: 'xyxy', xmin, ymin, xmax, ymax } as BoundingBox<F>;
+    case 'xywh':
+      return { format: 'xywh', xmin, ymin, w: xmax - xmin, h: ymax - ymin } as BoundingBox<F>;
+    case 'cxcywh':
+      return {
+        format: 'cxcywh',
+        cx: (xmin + xmax) / 2,
+        cy: (ymin + ymax) / 2,
+        w: xmax - xmin,
+        h: ymax - ymin,
+      } as BoundingBox<F>;
+  }
 }
 
 /**
@@ -159,7 +171,12 @@ export function splitTallQuad(ordered: readonly Point[], parts: number): Point[]
   for (let i = 0; i < parts; i++) {
     const t0 = i / parts;
     const t1 = (i + 1) / parts;
-    out.push([lerp(tl, bl, t0), lerp(tr, br, t0), lerp(tr, br, t1), lerp(tl, bl, t1)]);
+    out.push([
+      interpolatePoint(tl, bl, t0),
+      interpolatePoint(tr, br, t0),
+      interpolatePoint(tr, br, t1),
+      interpolatePoint(tl, bl, t1),
+    ]);
   }
   return out;
 }
@@ -179,7 +196,7 @@ export function boundingQuadOf(quads: readonly (readonly Point[])[]): Point[] {
       all.push(p);
     }
   }
-  const { xmin, ymin, xmax, ymax } = boundsOfPoints(all);
+  const { xmin, ymin, xmax, ymax } = boundsOfPoints(all, 'xyxy');
   return [
     { x: xmin, y: ymin },
     { x: xmax, y: ymin },
@@ -204,8 +221,8 @@ export function flattenQuad(corners: readonly Point[]): number[] {
 }
 
 /**
- * Builds oriented quads from a detector's flat output array — 10 numbers per box:
- * `x0,y0,..,x3,y3,score,angle`.
+ * Builds oriented quads from a detector's flat output array — 9 numbers per box:
+ * `x0,y0,..,x3,y3,score`.
  * @category Typescript API
  * @param flat The flat number array from a native detector decode.
  * @returns The parsed quads.
@@ -213,7 +230,7 @@ export function flattenQuad(corners: readonly Point[]): number[] {
 export function quadsFromFlat(flat: number[]): Quad[] {
   'worklet';
   const quads: Quad[] = [];
-  for (let i = 0; i < flat.length; i += 10) {
+  for (let i = 0; i < flat.length; i += 9) {
     quads.push({
       points: [
         { x: flat[i]!, y: flat[i + 1]! },
@@ -222,7 +239,6 @@ export function quadsFromFlat(flat: number[]): Quad[] {
         { x: flat[i + 6]!, y: flat[i + 7]! },
       ],
       score: flat[i + 8]!,
-      angle: flat[i + 9]!,
     });
   }
   return quads;
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
index 0c2dc3d283..c4e02dbf8c 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/detectors.ts
@@ -16,8 +16,7 @@ import { quadsFromFlat, type Quad } from '../../ops/quad';
  * new detector into the OCR pipeline by supplying a function of this type (the
  * built-ins below, or its own). MUST be a worklet.
  * @category Types
- * @param outputs The model's `detect_<S>` output tensors, in order (owned by the
- * pipeline — do not dispose).
+ * @param outputs The model's `detect_<S>` output tensors, in order
  * @param side The snapped square detector side `S` (input is `S × S` letterboxed).
  * @param charLevel Emit one box per glyph instead of grouped lines; strategies
  * without a char-level mode ignore it.
@@ -30,35 +29,29 @@ export type TextBoxExtractor = (
 ) => Quad[];
 
 /**
- * Threshold overrides for {@link makeCraftExtractBoxes}. Any omitted field keeps
- * the CRAFT default.
+ * Threshold overrides for {@link makeCraftExtractBoxes}. Omitted fields keep the
+ * CRAFT defaults: `textHeatmapThreshold` 0.4, `linkHeatmapThreshold` 0.4,
+ * `minBoxPeakScore` 0.7.
  * @category Types
  */
 export type CraftExtractorOptions = {
-  /** Region-heatmap binarization threshold. Defaults to 0.4. */
-  readonly textThreshold?: number;
-  /** Affinity-heatmap (glyph linking) threshold. Defaults to 0.4. */
-  readonly linkThreshold?: number;
-  /** Minimum component peak score to keep a box. Defaults to 0.7. */
-  readonly lowTextThreshold?: number;
+  readonly textHeatmapThreshold?: number;
+  readonly linkHeatmapThreshold?: number;
+  readonly minBoxPeakScore?: number;
 };
 
 /**
- * Threshold overrides for {@link makeDbnetExtractBoxes}. Any omitted field keeps
- * the DBNet default.
+ * Threshold overrides for {@link makeDbnetExtractBoxes}. Omitted fields keep the
+ * DBNet defaults: `binarizationThreshold` 0.3, `minBoxScore` 0.6, `unclipRatio`
+ * 1.5, `minBoxSidePx` 3, `maxContourCandidates` 1000.
  * @category Types
  */
 export type DbnetExtractorOptions = {
-  /** Probability-map binarization threshold. Defaults to 0.3. */
-  readonly binThreshold?: number;
-  /** Minimum mean in-contour probability to keep a box. Defaults to 0.6. */
-  readonly boxThreshold?: number;
-  /** Box expansion (unclip) ratio. Defaults to 1.5. */
+  readonly binarizationThreshold?: number;
+  readonly minBoxScore?: number;
   readonly unclipRatio?: number;
-  /** Minimum box side in pixels. Defaults to 3. */
-  readonly minBoxSide?: number;
-  /** Maximum contour candidates considered. Defaults to 1000. */
-  readonly maxCandidates?: number;
+  readonly minBoxSidePx?: number;
+  readonly maxContourCandidates?: number;
 };
 
 // CRAFT region+affinity heatmap thresholds — stable across models, the defaults.
@@ -74,7 +67,7 @@ const DBNET_MIN_BOX_SIDE = 3;
 const DBNET_MAX_CANDIDATES = 1000;
 
 /**
- * Builds a CRAFT {@link TextBoxExtractor} with custom thresholds: groups the
+ * Builds a CRAFT {@link TextBoxExtractor}. Groups the
  * half-resolution region+affinity heatmap (`outputs[0]` is the `[1,Hd,Wd,2]`
  * heatmap) into oriented text-line quads, or per-glyph boxes when `charLevel`.
  * For the standard thresholds use the ready-made {@link craftExtractBoxes}.
@@ -83,9 +76,9 @@ const DBNET_MAX_CANDIDATES = 1000;
  * @returns A {@link TextBoxExtractor} to assign to `OcrOptions.extractBoxes`.
  */
 export function makeCraftExtractBoxes(overrides?: CraftExtractorOptions): TextBoxExtractor {
-  const textThreshold = overrides?.textThreshold ?? CRAFT_TEXT_THRESHOLD;
-  const linkThreshold = overrides?.linkThreshold ?? CRAFT_LINK_THRESHOLD;
-  const lowTextThreshold = overrides?.lowTextThreshold ?? CRAFT_LOW_TEXT_THRESHOLD;
+  const textThreshold = overrides?.textHeatmapThreshold ?? CRAFT_TEXT_THRESHOLD;
+  const linkThreshold = overrides?.linkHeatmapThreshold ?? CRAFT_LINK_THRESHOLD;
+  const lowTextThreshold = overrides?.minBoxPeakScore ?? CRAFT_LOW_TEXT_THRESHOLD;
   return (outputs, side, charLevel) => {
     'worklet';
     // The half-resolution heatmap requires an even detector side; the pipeline is
@@ -107,7 +100,7 @@ export function makeCraftExtractBoxes(overrides?: CraftExtractorOptions): TextBo
 }
 
 /**
- * Builds a DBNet {@link TextBoxExtractor} with custom thresholds: thresholds and
+ * Builds a DBNet {@link TextBoxExtractor}. Thresholds and
  * unclips the probability map (`outputs[0]` is the `[1,1,H,W]` post-sigmoid prob
  * map) into oriented text quads. It decodes at full resolution with no char-level
  * mode, so the extractor uses neither `side` nor `charLevel`. For the standard
@@ -117,11 +110,11 @@ export function makeCraftExtractBoxes(overrides?: CraftExtractorOptions): TextBo
  * @returns A {@link TextBoxExtractor} to assign to `OcrOptions.extractBoxes`.
  */
 export function makeDbnetExtractBoxes(overrides?: DbnetExtractorOptions): TextBoxExtractor {
-  const binThreshold = overrides?.binThreshold ?? DBNET_BIN_THRESHOLD;
-  const boxThreshold = overrides?.boxThreshold ?? DBNET_BOX_THRESHOLD;
+  const binThreshold = overrides?.binarizationThreshold ?? DBNET_BIN_THRESHOLD;
+  const boxThreshold = overrides?.minBoxScore ?? DBNET_BOX_THRESHOLD;
   const unclipRatio = overrides?.unclipRatio ?? DBNET_UNCLIP_RATIO;
-  const minBoxSide = overrides?.minBoxSide ?? DBNET_MIN_BOX_SIDE;
-  const maxCandidates = overrides?.maxCandidates ?? DBNET_MAX_CANDIDATES;
+  const minBoxSide = overrides?.minBoxSidePx ?? DBNET_MIN_BOX_SIDE;
+  const maxCandidates = overrides?.maxContourCandidates ?? DBNET_MAX_CANDIDATES;
   return (outputs) => {
     'worklet';
     const flat = rnexecutorchJsi.cv.extractDbnetTextBoxes(outputs[0]!, {
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
deleted file mode 100644
index 8f4e3f6d1a..0000000000
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentHelpers.ts
+++ /dev/null
@@ -1,126 +0,0 @@
-import type { ImageBuffer } from '../../image';
-import { FORMAT_CHANNELS } from '../../ops/image';
-import type { BoundingBox } from '../../ops/boxes';
-import { boundsOfPoints } from '../../ops/quad';
-import type { OcrDetection } from '../ocr';
-
-/**
- * Crops an axis-aligned region out of an image as a plain pixel slice (same format
- * and layout). Used to feed a layout region to another model.
- * @category Typescript API
- * @param input The source image.
- * @param bbox The crop region, in `xyxy` pixels.
- * @returns The cropped image.
- */
-export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>): ImageBuffer {
-  'worklet';
-  const { data, width, height, format } = input;
-  const channels = FORMAT_CHANNELS[format];
-  const x0 = Math.max(0, Math.min(Math.round(bbox.xmin), width));
-  const y0 = Math.max(0, Math.min(Math.round(bbox.ymin), height));
-  const x1 = Math.max(0, Math.min(Math.round(bbox.xmax), width));
-  const y1 = Math.max(0, Math.min(Math.round(bbox.ymax), height));
-  const cropWidth = Math.max(1, x1 - x0);
-  const cropHeight = Math.max(1, y1 - y0);
-  const out = new Uint8Array(cropWidth * cropHeight * channels);
-  for (let y = 0; y < cropHeight; y++) {
-    const rowStart = ((y0 + y) * width + x0) * channels;
-    out.set(data.subarray(rowStart, rowStart + cropWidth * channels), y * cropWidth * channels);
-  }
-  return { data: out, width: cropWidth, height: cropHeight, format, layout: input.layout };
-}
-
-// Groups values into `k` ordered clusters by cutting the sorted values at their
-// (k-1) widest gaps, then returns each cluster's mean. Cutting at the widest gaps
-// (rather than at fixed intervals) keeps dense rows/columns together.
-function clusterCentersByGaps(values: readonly number[], k: number): number[] {
-  'worklet';
-  const sorted = [...values].sort((a, b) => a - b);
-  if (sorted.length <= k) {
-    return sorted;
-  }
-  // Rank the interior gaps (gap i sits between sorted[i-1] and sorted[i]) and take
-  // the k-1 widest as cut points, restored to ascending order.
-  const gaps = sorted.slice(1).map((value, i) => ({ at: i + 1, size: value - sorted[i]! }));
-  gaps.sort((a, b) => b.size - a.size);
-  const cuts = gaps
-    .slice(0, k - 1)
-    .map((gap) => gap.at)
-    .sort((a, b) => a - b);
-  // Average each [prev, cut) span into its center.
-  const centers: number[] = [];
-  let prev = 0;
-  for (const cut of [...cuts, sorted.length]) {
-    const group = sorted.slice(prev, cut);
-    centers.push(group.reduce((sum, value) => sum + value, 0) / group.length);
-    prev = cut;
-  }
-  return centers;
-}
-
-// Index of the center in `centers` nearest to `value`.
-function nearestIndex(value: number, centers: readonly number[]): number {
-  'worklet';
-  let best = 0;
-  for (let i = 1; i < centers.length; i++) {
-    if (Math.abs(value - centers[i]!) < Math.abs(value - centers[best]!)) {
-      best = i;
-    }
-  }
-  return best;
-}
-
-/**
- * Fills a table-structure HTML skeleton with a region's OCR lines. The grid size
- * comes from the skeleton (row count, and the widest row's cell count); each
- * line's box center is assigned to its nearest row and column cluster, so shared
- * column centers keep columns aligned. Falls back to a document-order fill when the
- * skeleton has no grid.
- *
- * Alignment is geometric only — dense rows can misplace a value, since the
- * skeleton carries no per-cell coordinates.
- * @category Typescript API
- * @param html The structure HTML skeleton (empty cells).
- * @param lines The region's OCR lines, with page-space quads.
- * @returns A `<table>` with each cell filled by its nearest-assigned text.
- */
-export function fillTableCells(html: string, lines: readonly OcrDetection[]): string {
-  'worklet';
-  const rowCount = (html.match(/<tr>/g) ?? []).length;
-  let colCount = 0;
-  const rowRegex = /<tr>([\s\S]*?)<\/tr>/g;
-  let row: RegExpExecArray | null;
-  while ((row = rowRegex.exec(html)) !== null) {
-    colCount = Math.max(colCount, (row[1]!.match(/<td/g) ?? []).length);
-  }
-  // No grid, or nothing to place: fill the skeleton cells in document order.
-  if (rowCount === 0 || colCount === 0 || lines.length === 0) {
-    let i = 0;
-    return html.replace(/<td([^>]*)><\/td>/g, (_match, attrs) => {
-      const text = i < lines.length ? lines[i]!.text : '';
-      i++;
-      return `<td${attrs}>${text}</td>`;
-    });
-  }
-
-  const centersX: number[] = [];
-  const centersY: number[] = [];
-  for (const line of lines) {
-    const box = boundsOfPoints(line.quad);
-    centersX.push((box.xmin + box.xmax) / 2);
-    centersY.push((box.ymin + box.ymax) / 2);
-  }
-  const rowCenters = clusterCentersByGaps(centersY, rowCount);
-  const colCenters = clusterCentersByGaps(centersX, colCount);
-  const grid: string[][] = Array.from({ length: rowCenters.length }, () =>
-    new Array<string>(colCenters.length).fill('')
-  );
-  for (let i = 0; i < lines.length; i++) {
-    const r = nearestIndex(centersY[i]!, rowCenters);
-    const c = nearestIndex(centersX[i]!, colCenters);
-    grid[r]![c] = `${grid[r]![c]!} ${lines[i]!.text}`.trim();
-  }
-  return `<table>${grid
-    .map((cells) => `<tr>${cells.map((text) => `<td>${text}</td>`).join('')}</tr>`)
-    .join('')}</table>`;
-}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
index 2fb3e65ad0..186f17239e 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentModels.ts
@@ -9,7 +9,10 @@ import type { ImageBuffer, ImageFormat } from '../../image';
 import { IMAGENET_NORM } from '../../../../constants';
 import { FORMAT_CHANNELS } from '../../ops/image';
 import { warpByGrid } from '../../ops/image';
+import type { BoundingBox } from '../../ops/boxes';
+import { boundsOfPoints } from '../../ops/quad';
 import { createImagePreprocessor } from '../preprocessing';
+import type { OcrDetection } from './ocr';
 
 /**
  * A detected page orientation: the clockwise rotation (rotate by its negation to
@@ -131,11 +134,9 @@ export async function createDocumentModels(
   const { modelPath, structureVocab, eosTokenId, maxSteps } = config;
   const model = await wrapAsync(loadModel, runtime)(modelPath);
 
-  // Contract validation + preprocessor/tensor construction below can throw (a
-  // missing method, shape/vocab mismatch, failed allocation). Everything built
-  // is pushed into `created` as it is created — one by one, so a mid-sequence
-  // failure can't strand its predecessors — and the catch disposes it all: a
-  // bad config must not leak native memory (mirrors createOcr).
+  // Everything built is pushed into `created` as it is created — one by one, so
+  // a mid-sequence failure can't strand its predecessors — and the catch
+  // disposes it all: a bad config must not leak native memory (mirrors createOcr).
   const created: { dispose: () => void }[] = [];
   try {
     // orientation: image -> class logits
@@ -327,3 +328,125 @@ export async function createDocumentModels(
     throw e;
   }
 }
+
+/**
+ * Crops an axis-aligned region out of an image as a plain pixel slice (same format
+ * and layout). Used to feed a layout region to another model.
+ * @category Typescript API
+ * @param input The source image.
+ * @param bbox The crop region, in `xyxy` pixels.
+ * @returns The cropped image.
+ */
+export function cropImageBuffer(input: ImageBuffer, bbox: BoundingBox<'xyxy'>): ImageBuffer {
+  'worklet';
+  const { data, width, height, format } = input;
+  const channels = FORMAT_CHANNELS[format];
+  const x0 = Math.max(0, Math.min(Math.round(bbox.xmin), width));
+  const y0 = Math.max(0, Math.min(Math.round(bbox.ymin), height));
+  const x1 = Math.max(0, Math.min(Math.round(bbox.xmax), width));
+  const y1 = Math.max(0, Math.min(Math.round(bbox.ymax), height));
+  const cropWidth = Math.max(1, x1 - x0);
+  const cropHeight = Math.max(1, y1 - y0);
+  const out = new Uint8Array(cropWidth * cropHeight * channels);
+  for (let y = 0; y < cropHeight; y++) {
+    const rowStart = ((y0 + y) * width + x0) * channels;
+    out.set(data.subarray(rowStart, rowStart + cropWidth * channels), y * cropWidth * channels);
+  }
+  return { data: out, width: cropWidth, height: cropHeight, format, layout: input.layout };
+}
+
+// 1-D clustering of cell-center coordinates into `k` table rows (or columns).
+// The sorted values are split at their k-1 widest gaps — each resulting run of
+// values is one row/column, represented by its mean coordinate. Splitting at the
+// widest gaps (instead of at fixed intervals) matches how table cells actually
+// distribute: values within a row are tightly packed while rows are separated by
+// clear gaps, so uneven row heights / column widths still cluster correctly.
+// Fewer than `k` values means every value is its own cluster.
+function clusterCentersByGaps(values: readonly number[], k: number): number[] {
+  'worklet';
+  const sorted = [...values].sort((a, b) => a - b);
+  if (sorted.length <= k) {
+    return sorted;
+  }
+  // Rank the interior gaps (gap i sits between sorted[i-1] and sorted[i]) and take
+  // the k-1 widest as cut points, restored to ascending order.
+  const gaps = sorted.slice(1).map((value, i) => ({ at: i + 1, size: value - sorted[i]! }));
+  gaps.sort((a, b) => b.size - a.size);
+  const cuts = gaps
+    .slice(0, k - 1)
+    .map((gap) => gap.at)
+    .sort((a, b) => a - b);
+  // Average each [prev, cut) span into its center.
+  const centers: number[] = [];
+  let prev = 0;
+  for (const cut of [...cuts, sorted.length]) {
+    const group = sorted.slice(prev, cut);
+    centers.push(group.reduce((sum, value) => sum + value, 0) / group.length);
+    prev = cut;
+  }
+  return centers;
+}
+
+/**
+ * Fills a table-structure HTML skeleton with a region's OCR lines. The grid size
+ * comes from the skeleton (row count, and the widest row's cell count); each
+ * line's box center is assigned to its nearest row and column cluster, so shared
+ * column centers keep columns aligned. Falls back to a document-order fill when the
+ * skeleton has no grid.
+ *
+ * Alignment is geometric only — dense rows can misplace a value, since the
+ * skeleton carries no per-cell coordinates.
+ * @category Typescript API
+ * @param html The structure HTML skeleton (empty cells).
+ * @param lines The region's OCR lines, with page-space quads.
+ * @returns A `<table>` with each cell filled by its nearest-assigned text.
+ */
+export function fillTableCells(html: string, lines: readonly OcrDetection[]): string {
+  'worklet';
+  const rowCount = (html.match(/<tr>/g) ?? []).length;
+  let colCount = 0;
+  const rowRegex = /<tr>([\s\S]*?)<\/tr>/g;
+  let row: RegExpExecArray | null;
+  while ((row = rowRegex.exec(html)) !== null) {
+    colCount = Math.max(colCount, (row[1]!.match(/<td/g) ?? []).length);
+  }
+  // No grid, or nothing to place: fill the skeleton cells in document order.
+  if (rowCount === 0 || colCount === 0 || lines.length === 0) {
+    let i = 0;
+    return html.replace(/<td([^>]*)><\/td>/g, (_match, attrs) => {
+      const text = i < lines.length ? lines[i]!.text : '';
+      i++;
+      return `<td${attrs}>${text}</td>`;
+    });
+  }
+
+  const centersX: number[] = [];
+  const centersY: number[] = [];
+  for (const line of lines) {
+    const box = boundsOfPoints(line.quad, 'xyxy');
+    centersX.push((box.xmin + box.xmax) / 2);
+    centersY.push((box.ymin + box.ymax) / 2);
+  }
+  const rowCenters = clusterCentersByGaps(centersY, rowCount);
+  const colCenters = clusterCentersByGaps(centersX, colCount);
+  const grid: string[][] = Array.from({ length: rowCenters.length }, () =>
+    new Array<string>(colCenters.length).fill('')
+  );
+  // Assign each line to the row/column whose cluster center is nearest.
+  for (let i = 0; i < lines.length; i++) {
+    const r = rowCenters.reduce(
+      (best, center, j) =>
+        Math.abs(centersY[i]! - center) < Math.abs(centersY[i]! - rowCenters[best]!) ? j : best,
+      0
+    );
+    const c = colCenters.reduce(
+      (best, center, j) =>
+        Math.abs(centersX[i]! - center) < Math.abs(centersX[i]! - colCenters[best]!) ? j : best,
+      0
+    );
+    grid[r]![c] = `${grid[r]![c]!} ${lines[i]!.text}`.trim();
+  }
+  return `<table>${grid
+    .map((cells) => `<tr>${cells.map((text) => `<td>${text}</td>`).join('')}</tr>`)
+    .join('')}</table>`;
+}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentOcr.ts
similarity index 93%
rename from packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentOcr.ts
index c565020d6c..6736432b48 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/documentOcr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/documentOcr.ts
@@ -1,21 +1,25 @@
 import type { WorkletRuntime } from 'react-native-worklets';
 
-import { tensor } from '../../../core/tensor';
-import { wrapAsync } from '../../../core/runtime';
-import type { ImageBuffer } from '../image';
-import type { Point } from '../ops/points';
-import type { BoundingBox } from '../ops/boxes';
-import { boundsOfPoints } from '../ops/quad';
-import { rotate, FORMAT_CHANNELS } from '../ops/image';
+import { tensor } from '../../../../core/tensor';
+import { wrapAsync } from '../../../../core/runtime';
+import type { ImageBuffer } from '../../image';
+import type { Point } from '../../ops/points';
+import type { BoundingBox } from '../../ops/boxes';
+import { boundsOfPoints } from '../../ops/quad';
+import { rotate, FORMAT_CHANNELS } from '../../ops/image';
 import { createOcr, type OcrModel, type OcrDetection } from './ocr';
 import {
   createObjectDetector,
   type ObjectDetectorModel,
   type ObjectDetection,
-} from './objectDetection';
-import { createDocumentModels, type DocumentModelsConfig } from './ocr/documentModels';
-import { orderByReadingOrder } from './ocr/ocrUtils';
-import { cropImageBuffer, fillTableCells } from './ocr/documentHelpers';
+} from '../objectDetection';
+import {
+  createDocumentModels,
+  cropImageBuffer,
+  fillTableCells,
+  type DocumentModelsConfig,
+} from './documentModels';
+import { orderByReadingOrder } from './ocrUtils';
 
 /**
  * One assembled document block: a layout region (or an ungrouped catch-all) with
@@ -226,7 +230,10 @@ export async function createDocumentOcr<L>(
           ? [
               makeBlock<L>(
                 'ungrouped',
-                { format: 'xyxy', ...boundsOfPoints(detections.flatMap((d) => d.quad as Point[])) },
+                boundsOfPoints(
+                  detections.flatMap((d) => d.quad as Point[]),
+                  'xyxy'
+                ),
                 1,
                 detections,
                 false
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocr.ts
similarity index 92%
rename from packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
rename to packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocr.ts
index 321dff054c..d1263873b1 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocr.ts
@@ -1,20 +1,15 @@
 import type { WorkletRuntime } from 'react-native-worklets';
 
-import { tensor, type Tensor } from '../../../core/tensor';
-import { loadModel } from '../../../core/model';
-import { wrapAsync } from '../../../core/runtime';
+import { tensor, type Tensor } from '../../../../core/tensor';
+import { loadModel } from '../../../../core/model';
+import { wrapAsync } from '../../../../core/runtime';
 
-import type { ImageBuffer } from '../image';
-import type { Point } from '../ops/points';
-import { FORMAT_CHANNELS, FORMAT_CONVERSION, cvtColor } from '../ops/image';
-import { orderQuad, quadSize, boundingQuadOf } from '../ops/quad';
-import type { TextBoxExtractor } from './ocr/detectors';
-import {
-  buildCharset,
-  orderByReadingOrder,
-  groupVerticalColumns,
-  type Buckets,
-} from './ocr/ocrUtils';
+import type { ImageBuffer } from '../../image';
+import type { Point } from '../../ops/points';
+import { FORMAT_CHANNELS, FORMAT_CONVERSION, cvtColor } from '../../ops/image';
+import { orderQuad, quadSize, boundingQuadOf } from '../../ops/quad';
+import type { TextBoxExtractor } from './detectors';
+import { orderByReadingOrder, groupVerticalColumns, type Buckets } from './ocrUtils';
 import {
   detectQuads,
   recognizeQuad,
@@ -27,11 +22,11 @@ import {
   type DetectContext,
   type RecContext,
   type VerticalContext,
-} from './ocr/pipeline';
+} from './pipeline';
 
-export type { Buckets } from './ocr/ocrUtils';
-export type { Quad } from '../ops/quad';
-export type { TextBoxExtractor } from './ocr/detectors';
+export type { Buckets } from './ocrUtils';
+export type { Quad } from '../../ops/quad';
+export type { TextBoxExtractor } from './detectors';
 
 /**
  * Configuration for the OCR pipeline: a model declares its input-size buckets, its
@@ -135,7 +130,7 @@ export type OcrDetection = {
   readonly confidence: number;
   /**
    * The oriented quad (TL,TR,BR,BL) in original image pixels. Derive the
-   * axis-aligned bounds with `boundsOfPoints(quad)` from `cv.ops.quad` if needed.
+   * axis-aligned bounds with `boundsOfPoints(quad, 'xyxy')` from `cv.ops.quad` if needed.
    */
   readonly quad: readonly Point[];
 };
@@ -241,7 +236,13 @@ export async function createOcr(
     if (recC !== 3) {
       throw new Error(`OCR: recognizer must take RGB (3 channels), but the model expects ${recC}.`);
     }
-    charset = buildCharset(ocrOpts.charset);
+    // CTC lookup: index 0 is the blank (ctcCollapse never decodes it), then the
+    // model's characters — a string splits into codepoints, an array is taken
+    // verbatim (preserving multi-codepoint entries like ligatures).
+    charset = [
+      '[blank]',
+      ...(typeof ocrOpts.charset === 'string' ? Array.from(ocrOpts.charset) : ocrOpts.charset),
+    ];
     if (charset.length !== rec.vocabSize) {
       throw new Error(
         `OCR: charset size (${charset.length}, incl. blank) must match recognizer output vocab (${rec.vocabSize}).`
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
index 4059e311a1..32c18131ad 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/ocrUtils.ts
@@ -20,10 +20,18 @@ export type Buckets = {
 // loses detail the model was trained to see.
 const BUCKET_SNAP_TOLERANCE = 0.1;
 
-// Selects the smallest bucket that fits `size`, but snaps down to the next-lower
-// bucket when `size` exceeds it by no more than BUCKET_SNAP_TOLERANCE. Clamps to
-// the largest bucket for oversized inputs.
-function snapBucket(size: number, buckets: readonly number[]): number {
+/**
+ * Selects the smallest bucket that fits `size`, but snaps down to the next-lower
+ * bucket when `size` exceeds it by no more than the snap tolerance; oversized
+ * inputs clamp to the largest bucket. Detector callers pass the image's longest
+ * side (selecting `detect_<S>`); recognizer callers pass the desired crop
+ * content width (selecting `recognize_<W>`).
+ * @category Typescript API
+ * @param size The size to fit, in pixels.
+ * @param buckets The model's ascending bucket sizes.
+ * @returns The selected bucket size.
+ */
+export function snapBucket(size: number, buckets: readonly number[]): number {
   'worklet';
   for (let i = 0; i < buckets.length; i++) {
     if (buckets[i]! >= size) {
@@ -34,35 +42,6 @@ function snapBucket(size: number, buckets: readonly number[]): number {
   return buckets[buckets.length - 1]!;
 }
 
-/**
- * Selects the detector bucket for an image from its longest side.
- * @category Typescript API
- * @param imageWidth The image width in pixels.
- * @param imageHeight The image height in pixels.
- * @param buckets The ascending detector side buckets.
- * @returns The selected square side `S` (invoke `detect_<S>`).
- */
-export function snapDetectBucket(
-  imageWidth: number,
-  imageHeight: number,
-  buckets: readonly number[]
-): number {
-  'worklet';
-  return snapBucket(Math.max(imageWidth, imageHeight), buckets);
-}
-
-/**
- * Selects the recognizer width bucket for a desired crop content width.
- * @category Typescript API
- * @param desiredWidth The crop content width at the recognizer height.
- * @param buckets The ascending recognizer width buckets.
- * @returns The selected width `W` (invoke `recognize_<W>`).
- */
-export function snapRecognizeBucket(desiredWidth: number, buckets: readonly number[]): number {
-  'worklet';
-  return snapBucket(desiredWidth, buckets);
-}
-
 /**
  * Computes the content width (px) of a recognizer crop: the region resized to the
  * recognizer height keeping its aspect ratio, clamped to the bucket width.
@@ -90,18 +69,25 @@ export function contentWidthFor(
 const COLUMN_GAP_FRACTION = 0.06;
 const LINE_OVERLAP_FRACTION = 0.3;
 
-// Returns the indices of `quads` in human reading order. Column gutters are found
-// by an x-coverage sweep (a band no box crosses, wider than COLUMN_GAP_FRACTION of
-// the content width, splits columns); within each column boxes are grouped into
-// lines by vertical overlap, lines ordered top-to-bottom, boxes within a line
-// left-to-right, and columns read left-to-right.
-function readingOrder(quads: readonly (readonly Point[])[]): number[] {
+/**
+ * Reorders items carrying a `quad` into human reading order: multi-column inputs
+ * read column-by-column, single-column inputs line-by-line, and boxes within a
+ * line left-to-right. Detectors emit boxes in an arbitrary order, so detections
+ * and assembled blocks are ordered through this.
+ * @category Typescript API
+ * @param items The items to reorder, each carrying a `quad`.
+ * @returns The items in reading order.
+ */
+export function orderByReadingOrder<T extends { quad: readonly Point[] }>(items: T[]): T[] {
   'worklet';
-  const count = quads.length;
+  const count = items.length;
   if (count <= 1) {
-    return count === 1 ? [0] : [];
+    return items;
   }
-  const boxes = quads.map((q) => boundsOfPoints(q));
+
+  // 1. Axis-aligned bounds per quad, plus the content x-range: a column gutter
+  //    must be at least COLUMN_GAP_FRACTION of that range to count.
+  const boxes = items.map((it) => boundsOfPoints(it.quad, 'xyxy'));
   let minX = Infinity;
   let maxX = -Infinity;
   for (const box of boxes) {
@@ -110,8 +96,9 @@ function readingOrder(quads: readonly (readonly Point[])[]): number[] {
   }
   const minGap = COLUMN_GAP_FRACTION * Math.max(1, maxX - minX);
 
-  // Sweep the box x-edges; an interior span with zero coverage wider than minGap
-  // is a column gutter, cut at its midpoint.
+  // 2. Find column gutters with an x-coverage sweep over the box edges: while
+  //    inside any box the coverage counter is > 0; a zero-coverage span wider
+  //    than minGap is a gutter, and its midpoint becomes a column cut.
   const edges: { x: number; delta: number }[] = [];
   for (const box of boxes) {
     edges.push({ x: box.xmin, delta: 1 });
@@ -132,7 +119,8 @@ function readingOrder(quads: readonly (readonly Point[])[]): number[] {
     }
   }
 
-  // Assign each box to a column by its center-x relative to the (ascending) cuts.
+  // 3. Assign each box to a column: count how many (ascending) cuts its
+  //    center-x lies to the right of.
   const columns: number[][] = Array.from({ length: cuts.length + 1 }, () => []);
   for (let i = 0; i < count; i++) {
     const centerX = (boxes[i]!.xmin + boxes[i]!.xmax) / 2;
@@ -143,6 +131,9 @@ function readingOrder(quads: readonly (readonly Point[])[]): number[] {
     columns[column]!.push(i);
   }
 
+  // 4. Within each (left-to-right) column: group boxes into lines by vertical
+  //    overlap (≥ LINE_OVERLAP_FRACTION of the shorter height joins a line),
+  //    order lines top-to-bottom and boxes within a line left-to-right.
   const order: number[] = [];
   for (const column of columns) {
     column.sort((a, b) => boxes[a]!.ymin - boxes[b]!.ymin);
@@ -173,24 +164,7 @@ function readingOrder(quads: readonly (readonly Point[])[]): number[] {
       order.push(...line.items);
     }
   }
-  return order;
-}
-
-/**
- * Reorders items carrying a `quad` into human reading order: multi-column inputs
- * read column-by-column, single-column inputs line-by-line, and boxes within a
- * line left-to-right. Detectors emit boxes in an arbitrary order, so detections
- * and assembled blocks are ordered through this.
- * @category Typescript API
- * @param items The items to reorder, each carrying a `quad`.
- * @returns The items in reading order.
- */
-export function orderByReadingOrder<T extends { quad: readonly Point[] }>(items: T[]): T[] {
-  'worklet';
-  if (items.length <= 1) {
-    return items;
-  }
-  return readingOrder(items.map((it) => it.quad)).map((i) => items[i]!);
+  return order.map((i) => items[i]!);
 }
 
 // A box wider than this multiple of its height is a horizontal line, never a
@@ -227,7 +201,7 @@ export function groupVerticalColumns(quads: readonly (readonly Point[])[]): {
   const candidates: Candidate[] = [];
   const singles: Point[][] = [];
   for (const q of quads) {
-    const { xmin, ymin, xmax, ymax } = boundsOfPoints(q);
+    const { xmin, ymin, xmax, ymax } = boundsOfPoints(q, 'xyxy');
     const width = xmax - xmin;
     const height = ymax - ymin;
     if (width > height * COLUMN_GLYPH_ASPECT) {
@@ -308,20 +282,3 @@ export function ctcCollapse(
   }
   return { text, confidence: count === 0 ? 0 : sum / count };
 }
-
-/**
- * Builds a CTC charset lookup: `numSpecials` reserved tokens (the CTC blank) are
- * prepended, then the characters follow — a string is split into codepoints, an
- * array is taken verbatim (preserving multi-codepoint entries) — so
- * `charset[index]` decodes argmax `index`.
- * @category Typescript API
- * @param charset The model's ordered character set.
- * @param numSpecials Number of reserved low indices (default 1 = CTC blank).
- * @returns The charset lookup array, `numSpecials` reserved slots at the front.
- */
-export function buildCharset(charset: string | readonly string[], numSpecials = 1): string[] {
-  'worklet';
-  const reserved = Array.from({ length: numSpecials }, (_unused, i) => `[reserved${i}]`);
-  const chars = typeof charset === 'string' ? Array.from(charset) : charset;
-  return [...reserved, ...chars];
-}
diff --git a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
index 316c1fd105..5397b41dd0 100644
--- a/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
+++ b/packages/react-native-executorch/src/extensions/cv/tasks/ocr/pipeline.ts
@@ -24,7 +24,7 @@ import {
 } from '../../ops/image';
 import { mapQuadToImage, orderQuad, quadSize, flattenQuad, splitTallQuad } from '../../ops/quad';
 import type { TextBoxExtractor } from './detectors';
-import { contentWidthFor, ctcCollapse, snapDetectBucket, snapRecognizeBucket } from './ocrUtils';
+import { contentWidthFor, ctcCollapse, snapBucket } from './ocrUtils';
 
 // The detector consumes raw RGB scaled to [0,1]; its mean/std normalization is
 // baked into the model, so the client only divides by 255.
@@ -119,7 +119,7 @@ export function detectQuads(
   charLevel = false
 ): Point[][] {
   'worklet';
-  const detS = snapDetectBucket(width, height, ctx.detBuckets);
+  const detS = snapBucket(Math.max(width, height), ctx.detBuckets);
   const detSet = ctx.detSets.get(detS)!;
   // Only the source resize depends on the run's channel count; the rest is cached.
   const tDetResize = tensor('uint8', [detS, detS, ctx.numChannels]);
@@ -173,7 +173,7 @@ export function recognizeQuad(
   const size = quadSize(corners);
   const maxRec = ctx.recBuckets[ctx.recBuckets.length - 1]!;
   const desiredW = contentWidthFor(size.width, size.height, ctx.recH, maxRec);
-  const bucketW = snapRecognizeBucket(desiredW, ctx.recBuckets);
+  const bucketW = snapBucket(desiredW, ctx.recBuckets);
   const recSet = ctx.recSetByWidth.get(bucketW)!;
   warpQuad(src, recSet.tCanvas, flattenQuad(corners), {
     contentWidth: Math.min(desiredW, bucketW),
@@ -220,7 +220,7 @@ export function recognizeGlyphStrip(
   if (cells.length === 0) {
     return null;
   }
-  const bucketW = snapRecognizeBucket(totalW, recCtx.recBuckets);
+  const bucketW = snapBucket(totalW, recCtx.recBuckets);
   const recSet = recCtx.recSetByWidth.get(bucketW)!;
   // Warp each cell into the canvas at its x-offset; the first warp clears + pads
   // the whole canvas, the rest compose in with `clear: false`.
diff --git a/packages/react-native-executorch/src/hooks/useDocumentOcr.ts b/packages/react-native-executorch/src/hooks/useDocumentOcr.ts
index beb1ecac91..810e2ff4cd 100644
--- a/packages/react-native-executorch/src/hooks/useDocumentOcr.ts
+++ b/packages/react-native-executorch/src/hooks/useDocumentOcr.ts
@@ -1,6 +1,6 @@
 import { useModel } from './useModel';
 import { useResourceDownload } from './useResourceDownload';
-import { createDocumentOcr, type DocumentOcrModel } from '../extensions/cv/tasks/documentOcr';
+import { createDocumentOcr, type DocumentOcrModel } from '../extensions/cv/tasks/ocr/documentOcr';
 
 // Swap a model spec's hosted `modelPath` for its downloaded local path. Returns
 // undefined when the spec is absent (an optional model) or its path hasn't
diff --git a/packages/react-native-executorch/src/hooks/useOcr.ts b/packages/react-native-executorch/src/hooks/useOcr.ts
index e8b9dc3962..d234502211 100644
--- a/packages/react-native-executorch/src/hooks/useOcr.ts
+++ b/packages/react-native-executorch/src/hooks/useOcr.ts
@@ -1,4 +1,4 @@
-import { createOcr, type OcrModel } from '../extensions/cv/tasks/ocr';
+import { createOcr, type OcrModel } from '../extensions/cv/tasks/ocr/ocr';
 import { useResourceDownload } from './useResourceDownload';
 import { useModel } from './useModel';
 
diff --git a/packages/react-native-executorch/src/index.ts b/packages/react-native-executorch/src/index.ts
index afcafa47f3..23f4e471c3 100644
--- a/packages/react-native-executorch/src/index.ts
+++ b/packages/react-native-executorch/src/index.ts
@@ -23,10 +23,10 @@ export * from './extensions/cv/tasks/instanceSegmentation';
 export * from './extensions/cv/tasks/keypointDetection';
 export * from './extensions/cv/tasks/objectDetection';
 export * from './extensions/nlp/tasks/tokenization';
-export * from './extensions/cv/tasks/ocr';
+export * from './extensions/cv/tasks/ocr/ocr';
 export * from './extensions/cv/tasks/ocr/detectors';
 export type { DocumentModelsConfig } from './extensions/cv/tasks/ocr/documentModels';
-export * from './extensions/cv/tasks/documentOcr';
+export * from './extensions/cv/tasks/ocr/documentOcr';
 
 // Core primitives — for library builders and power users
 export { tensor } from './core/tensor';
diff --git a/packages/react-native-executorch/src/models.ts b/packages/react-native-executorch/src/models.ts
index 7a79f13329..567a667e13 100644
--- a/packages/react-native-executorch/src/models.ts
+++ b/packages/react-native-executorch/src/models.ts
@@ -4,7 +4,7 @@ import type { StyleTransferModel } from './extensions/cv/tasks/styleTransfer';
 import type { SemanticSegmentationModel } from './extensions/cv/tasks/semanticSegmentation';
 import type { KeypointDetectorModel } from './extensions/cv/tasks/keypointDetection';
 import type { InstanceSegmenterModel } from './extensions/cv/tasks/instanceSegmentation';
-import type { OcrModel, OcrOptions } from './extensions/cv/tasks/ocr';
+import type { OcrModel, OcrOptions } from './extensions/cv/tasks/ocr/ocr';
 import { craftExtractBoxes, dbnetExtractBoxes } from './extensions/cv/tasks/ocr/detectors';
 import type { DocumentModelsConfig } from './extensions/cv/tasks/ocr/documentModels';
 import {