@unlikeotherai/faces

Browser recognition package for Faces.

This package owns the web-side recognition loop:

camera frame source -> Web Worker client -> model embedder -> embedding -> match

It is framework-agnostic. A consuming POS app decides when recognition runs and what a successful match does.

Usage

import {
  createFacesWebRecognition,
  createFacesRecognitionWorker,
  createRecognitionWorkerClient,
} from '@unlikeotherai/faces';

const worker = createFacesRecognitionWorker();
const workerClient = createRecognitionWorkerClient(worker);
await workerClient.initialize({
  moduleUrl: new URL('./modelEmbedder.ts', import.meta.url).href,
  allowedModuleOrigins: [location.origin],
  expectedEmbeddingDimensions: 128,
  wasmUrl: '/faces/model.wasm',
  modelUrl: '/faces/model.bin',
});

const recognition = createFacesWebRecognition({ workerClient });

recognition.replaceWorkerEmbeddings([
  {
    workerId: 'worker-1',
    workerName: 'Alice',
    embeddings: [new Float32Array([1, 0, 0])],
  },
]);

const unsubscribe = recognition.onFaceRecognized((match) => {
  console.log(match.workerId, match.score);
});

Submitted frames are owned by the recognition layer. Do not reuse a VideoFrame, ImageBitmap, or ImageData after submitting it.

The model embedder module is supplied by the application or model export layer. It must expose createEmbedder(config) and return an object with embedFrame. Model modules default to same-origin loading; cross-origin modules must be explicitly allowlisted.

Model weights are not included in this package.

Build

pnpm install
pnpm run build

The published package serves files from dist/.