Client-side face detection with TensorFlow.js

How it runs

The model downloads once and executes locally against WebGL. No frame, no upload, and no detection result leaves the device, which is less a privacy stance than a description of where the arithmetic happens: there is no server in this page to send anything to.

BlazeFace returns boxes and six landmarks per face and is built for speed over precision, so expect it to hold up on frontal faces at close range and to drop profiles, small faces, and heavy occlusion. Results vary by device and browser with the GPU underneath.

Known gaps

• No camera selector, so multi-camera devices get whichever one the browser picks first.
• Still on the standalone blazeface package rather than the newer face-detection module, which wraps the same model behind a maintained API.

For the same problem solved natively, with numbers: Apple Vision vs YOLOv11n.