React Native ML Kit — Face Recognition

@nitro-mlkit/face-recognition · on-device Google ML Kit via Nitro Modules — JSI, no bridge.

⚠️ Beta (0.1.0-beta.x). Android verified on-device; iOS impl pending (see Platform status). You provide the embedding model.

On-device face recognition for React Native, built with Nitro Modules (JSI, no bridge).

Google ML Kit finds & crops faces; a TensorFlow Lite face-embedding model (you supply it — e.g. MobileFaceNet, ~5 MB) turns each face into a vector, and cosine similarity against an in-memory registry answers "who is this?". Built for flows like a party game: register players by selfie, then scan a gallery to find which photos each player appears in.

Why bring your own model? ML Kit only does face detection, not recognition/embeddings. Face-embedding models vary in size/quality/license, so this package ships none (keeping it tiny) and loads one at runtime. Input and output tensor shapes are read from your model, so 112×112 / 128-d / 192-d MobileFaceNet variants all work.

Installation

npm install @nitro-mlkit/face-recognition react-native-nitro-modules

No config plugin (autolinked Expo module). Install and npx expo prebuild. Not available in Expo Go.

Usage

import { NitroRecognizer } from "@nitro-mlkit/face-recognition";

// 1. Provide a face-embedding model once (cached on disk).
await NitroRecognizer.downloadModel("https://your-host/mobilefacenet.tflite");
// …or load one you've bundled / downloaded yourself:
// await NitroRecognizer.loadModel("file:///…/mobilefacenet.tflite");

// 2. Register people (e.g. from selfies).
await NitroRecognizer.registerPerson("marcos", "Marcos", selfieUri);
await NitroRecognizer.registerPerson("lucia", "Lucía", selfieUri2);
await NitroRecognizer.addReference("marcos", anotherMarcosPhoto); // improves accuracy

// 3. Scan photos — one native batch call.
const results = await NitroRecognizer.findPeopleInPhotos(galleryUris, {
  concurrency: 4,
  minSimilarity: 0.6,
});
const marcosPhotos = results.filter((r) =>
  r.people.some((p) => p.person.id === "marcos"),
);

// Lower-level:
const emb = await NitroRecognizer.extractEmbedding(faceUri); // { vector: number[] }
const sim = NitroRecognizer.compare(embA, embB);             // cosine, 0..1

NitroRecognizer.clearRegistry(); // e.g. end of game session

Choosing a model

Any TFLite face-embedding model that takes an RGB face crop and outputs a fixed-length vector works (pixels are normalised as (px − 127.5) / 128). MobileFaceNet (Apache-2.0, ~5 MB, 112×112 → 192-d) is a good default. Host it yourself (e.g. a GitHub release asset or your CDN) and pass the URL to downloadModel(). A curated FaceModel enum of self-hosted, license-clear models is planned for a later release.

Platform status

| Platform | Min | Status | | -------- | --- | ------ | | Android | API 26+ | ✅ Verified on-device (Pixel 9, API 36): register + find + compare | | iOS | 15.5+ | 🔨 Swift impl pending (GoogleMLKit + TensorFlowLite) |

Part of nitro-mlkit

The full ML Kit suite on Nitro — see the other @nitro-mlkit/* packages.

License

MIT © Gonzalo Polo · (bring-your-own model keeps its own license)