react-native-vision-camera (v5)

VisionCamera v5 is the maintained and latest version of react-native-vision-camera. It is a full Nitro Modules rewrite with a new Constraints API, Output-based architecture, in-memory Photo, and a hard break from the v4 format/prop model. Almost every v4 surface is gone or renamed — treat v5 as a new API, not an incremental upgrade.

This skill is a router. Read this file first, then load the reference that matches the task. Every reference is self-contained — do not load more than you need.

When to load which reference

• Production low-latency GPU, ML, CV, or frame-coupled overlay work: use the separate react-native-vision-camera-realtime skill. Load the basic frame-processing reference too only when setup or API fundamentals are also needed.
• New install, getting a Camera on screen, permissions, minimum boilerplate → references/quickstart-v5.md
• Porting a v4 codebase, understanding what changed → references/migration-v4-to-v5.md (load this FIRST when the user mentions v4, takePhoto, useCameraFormat, format prop, photo/video boolean props, or CodeScanner in core)
• Porting a whole v4 screen — want a complete before/after file to transplant → references/migration-templates.md (full copy-paste templates: photo screen, video screen, frame-processor+ML, barcode scanner, pro camera)
• Choosing/attaching outputs, fps/HDR/resolution via constraints, session lifecycle → references/outputs-and-constraints.md
• Frame Processors, worklets, async frame work, pixel formats, writing a native plugin → references/frame-processors.md (load this when user says "frame processor", "worklet", "ML on frames", "Nitro plugin", "vision-camera-plugin-*")
• Capturing photos (incl. callbacks, RAW, HDR, preview image), recording video, Recorder lifecycle, manual AE/AF/AWB, exposure bias, zoom, focus → references/capture-and-controls.md
• Depth streaming, multi-cam, Skia preview, GPU resizer for ML, barcode scanner package, GPS location metadata, custom native outputs → references/advanced-features.md

When in doubt, load references/migration-v4-to-v5.md — it covers the shape of the new API by contrasting it with v4 and is the fastest orientation.

Non-negotiable rules for v5 code

These are the rules that catch people who "know" v4. Apply them without asking:

1. Install the Nitro peers. react-native-nitro-modules and react-native-nitro-image are required peer deps. Frame processors additionally require react-native-vision-camera-worklets AND react-native-worklets (Software Mansion's — not -core). Worklets - https://docs.swmansion.com/react-native-worklets/docs/
2. outputs={[...]} replaces photo / video / frameProcessor / codeScanner props. Create outputs with usePhotoOutput, useVideoOutput, useFrameOutput, useDepthOutput, useObjectOutput (or useBarcodeScannerOutput from the barcode package) and pass them in an array. Capture methods (capturePhoto, createRecorder) live on the Output, not the Camera ref.
3. There is no format prop and no useCameraFormat. Use constraints={[...]} — array order = priority, descending. The Camera negotiates the closest supported config automatically, so an impossible constraint like { fps: 99999 } never throws.
4. takePhoto() does not exist. Use photoOutput.capturePhoto(settings, callbacks) for in-memory Photo, or photoOutput.capturePhotoToFile(...) for a file path. The default path is in-memory — do not write temp files unless explicitly asked.
5. Frame Processor plugins must be Nitro Modules. The v4 FrameProcessorPlugin base class, VISION_EXPORT_SWIFT_FRAME_PROCESSOR macro, and VisionCameraProxy.addFrameProcessorPlugin are gone. A v5 plugin is a HybridObject with a typed Nitro spec. See references/frame-processors.md.
6. Every Frame (and Depth) MUST be .dispose()d. The buffer pool is bounded; leaking a frame stalls the pipeline. Wrap work in try { ... } finally { frame.dispose() }. When offloading via asyncRunner.runAsync(...), dispose inside the async callback if it returned true, and dispose immediately in the else branch when it returned false.
7. CodeScanner is not in core. react-native-vision-camera-barcode-scanner is a separate package, MLKit-based on both platforms. For iOS-only object detection (QR, faces, bodies via native AVFoundation metadata, no ML dep), use useObjectOutput from core.
8. Keep the Camera mounted; toggle isActive. Remounting tears down the session. Integrate with useIsFocused() from react-navigation so the session goes Idle → Ready while not on screen, and keeps preferences warm for fast resume.
9. Frame output pixelFormat defaults to 'native' (zero-copy), NOT 'yuv'. 'native' streams in the session's negotiated nativePixelFormat with zero conversions (it may resolve to a YUV, RGB, RAW, or 'private' format; verify the actual one via frame.pixelFormat). 'yuv' picks the YUV format closest to native and is the best general-purpose CPU-accessible choice (MLKit/OpenCV/Skia); 'rgb' forces a YUV-to-RGB conversion with about 2.6 times more bandwidth, so use it only when a consumer hard-requires RGB. useDepthOutput has no pixelFormat option. For ML consumers that require CPU-visible RGB or tensor input, prefer react-native-vision-camera-resizer over paying a per-frame RGB conversion in the Camera pipeline.

<!-- source: useFrameOutput.ts:123 (`pixelFormat = 'native'` default); VideoPixelFormat.ts:52-62; CameraFrameOutput.nitro.ts:71,84-86 ("recommended to use 'native' ... zero-copy GPU-only path"); useDepthOutput.ts:70-77 (options have no pixelFormat) -->

10. Do not hand-clamp FPS/resolution with Math.min/Math.max. That was a v4 workaround. In v5 the Constraints API negotiates internally — express intent and let the Camera pick.
11. Worklets mutate Reanimated SharedValues directly in v5. This is suitable for ordinary asynchronous UI or animation state. For frame-locked overlays, use the separate real-time skill and draw from the same frame with Skia or WebGPU.

Operating rules for this skill

• Never invent v4→v5 API shapes. If a v4 API has no documented v5 equivalent in the references, say so and link to the v5 docs — do not guess.
• Do not add documentation files (README, CHANGELOG) unless the user asks.
• Assume the user is on v5 unless they show v4 code. If they show v4 code, load references/migration-v4-to-v5.md before writing anything.
• When writing a new Camera example, default to the hook-based declarative form (useCameraPermission + useCameraDevice + usePhotoOutput + <Camera />). Use the imperative VisionCamera.createCameraSession(...) API only when the user asks for multi-cam or full programmatic control.
• For a basic ML path whose consumer needs CPU-visible input, recommend react-native-vision-camera-resizer over the v4-era vision-camera-resize-plugin. Route latency-critical ML, GPU inference, and live overlays to react-native-vision-camera-realtime instead of prescribing the Resizer universally.

<!-- source: react-native-vision-camera-resizer/src/specs/GPUFrame.nitro.ts + Resizer.nitro.ts (GPU resize→GPUFrame). The often-quoted "~5×" figure is a blog claim, not in source, so it is omitted here. -->

• Verify peer dependency installs. A user reporting a native crash after install 95% of the time has missed react-native-nitro-modules, react-native-nitro-image, or (for frame processors) react-native-worklets + react-native-vision-camera-worklets.

Authoritative links

• Docs: https://visioncamera.margelo.com
• llms.txt index: https://visioncamera.margelo.com/llms.txt
• V5 release notes (includes migration snippets): gh api repos/mrousavy/react-native-vision-camera/releases/tags/v5.0.0
• Blog announcement: https://blog.margelo.com/whats-new-in-visioncamera-v5
• Main repo: https://github.com/mrousavy/react-native-vision-camera
• V4 snapshot (archived docs): https://visioncamera4.margelo.com