Features

• Cutouts — Isolate the subject (removeBackground). Trim, export PNG/JPEG, optional mask.
• Labels — Ranked classifications (classifyImage). Offline on both platforms.
• OCR — Read text (readText). Latin / CJK / Devanagari on Android; Vision on iOS 18+.
• Compose — One decode, many results (analyzeImage). Segment + classify + OCR together.
• Native — Nitro HybridObjects keep large results on the native side. Call dispose() when done.

Install

npm install react-native-nitro-vision-kit react-native-nitro-modules
cd ios && pod install

| | Minimum | | --- | --- | | React Native | 0.75 | | react-native-nitro-modules | 0.36.0 | | iOS | 13 · cutouts 17+ · OCR 18+ | | Android | API 24 |

Prefer the New Architecture — Nitro is built for it.

Quick start

import { VisionKit } from 'react-native-nitro-vision-kit'

Cutout

if (!VisionKit.capabilities.supportsBackgroundRemoval) {
  throw new Error(VisionKit.capabilities.backgroundRemovalUnavailableReason)
}

const cutout = await VisionKit.removeBackground(imagePath, { trim: true })
const path = await cutout.saveToTemporaryFile('png', 100)
cutout.dispose()

Labels

const labels = await VisionKit.classifyImage(imagePath, {
  maxResults: 5,
  minConfidence: 0.5,
})

OCR

const ocr = await VisionKit.readText(imagePath)
console.log(ocr.text)
ocr.dispose()

Compose (one decode)

const result = await VisionKit.analyzeImage(imagePath, {
  removeBackground: { trim: true },
  classify: { maxResults: 5 },
  readText: {},
})

result.segmentation?.dispose()
result.text?.dispose()

[!NOTE] removeBackground and readText return HybridObjects. Export or read what you need, then call dispose().

API

Capabilities

VisionKit.capabilities

| Field | Meaning | | --- | --- | | supportsBackgroundRemoval | Cutouts available | | backgroundRemovalUnavailableReason | Why cutouts are off | | supportsImageClassification | Labels available | | supportsTextRecognition | OCR available | | supportedTextLanguages | Language tags you can request |

removeBackground(path, options?)

Returns a cutout HybridObject.

| Platform | Requirement | | --- | --- | | iOS | 17+ | | Android | API 24+, Play Services, ML Kit subject segmentation (beta) |

Export: saveToTemporaryFile(format, quality) · toArrayBuffer() · toMaskBuffer()

| Option | Default | Meaning | | --- | --- | --- | | trim | true | Crop to the subject | | maxPixels | 6_000_000 | Cap when loading the image | | retainMask | false | Keep mask for toMaskBuffer() |

Result: width, height, bounds (VisionRect, 0–1), pixelBounds, foregroundCoverage, centroid, instanceCount, hasMask, sourceWidth, sourceHeight, trimOrigin

classifyImage(path, options?)

Returns { label, confidence, index }[], highest confidence first.

| Platform | Notes | | --- | --- | | iOS | 13+ | | Android | Model ships with the library — offline |

| Option | Default | Meaning | | --- | --- | --- | | maxResults | 0 | 0 keeps all above the score floor | | minConfidence | 0.5 | Lowest score to keep | | region | full image | VisionRect (0–1) |

readText(path, options?)

Returns a text HybridObject. Prefer text and blockAt(i) — blocks copies everything into JS.

| Platform | Notes | | --- | --- | | iOS | 18+ (RecognizeTextRequest) | | Android | Play Services: Latin, Chinese, Japanese, Korean, Devanagari |

Image load cap: 4M pixels (both platforms). On Android, longest side is also capped at 2048.

| Option | Default | Platform | | --- | --- | --- | | languages | auto / Latin | Both | | recognitionLevel | accurate | iOS | | region | full image | Both (VisionRect) | | minTextHeightFraction | unset | Both | | usesLanguageCorrection | true | iOS | | customWords | unset | iOS | | maxCandidates | 1 | iOS (1–10) |

Android: languages selects script models. Non-Latin models also read Latin. Multiple non-Latin scripts can run together. A block may contain many lines.

iOS: Each block is one line.

analyzeImage(path, options)

One decode. Pass at least one of removeBackground, classify, or readText.

• No subject found → segmentation is omitted; labels and text still run
• If classify/OCR omit region and a cutout ran → subject bounds are used

Platform notes

Model download (Android)

| Method | First use | | --- | --- | | classifyImage | Offline — model is packaged | | removeBackground / readText | Downloads a Play Services model once |

Online: wait up to ~2 minutes for the first download, then offline. Offline with no model → fails immediately. iOS models ship with the system — no download step.

Paths

| Input | iOS | Android | | --- | --- | --- | | Absolute path | yes | yes | | file:// | yes | yes | | content:// | no | yes |

[!IMPORTANT] Only pass paths created by your app. Native code opens the path and returns pixels/text to JavaScript.

Example app

cd example
npm install
cd ios && bundle install && bundle exec pod install && cd ..
npm run ios    # or: npm run android

After a cutout, Keep saves to Photos.

License

MIT · Security · Changelog