FaceX 01 / 25
In-browser face detection & 576-point 3D mesh
Face detection, 98-point landmarks and a dense 576-point 3D face mesh running entirely in the browser via WebAssembly.
Problem
Face analysis usually means shipping camera frames to a server or a heavy native SDK. That costs latency, money and privacy.
What it does
- Detects faces on live camera frames
- Predicts 98 landmarks (WFLW-style) per face
- Builds a 576-point 3D mesh (distilled from MediaPipe) for AR/analytics
- Runs with zero server: camera frames never leave the device
How it works
- Camera frame → YuNet-style FCOS face detector (401 KB, trained on WIDER FACE)
- Face crop → 98-point landmark model (1.1 MB, WFLW)
- Landmarks → 576-point 3D mesh model (5.6 MB, MediaPipe distillation)
- All models compiled into WebAssembly; weights shipped as AES-256-GCM ciphertext and decrypted in the browser via WebCrypto
Stack & models
CWebAssembly (SIMD)WebCryptoModels trained in-house
Numbers
- Detector 401 KB · landmarks 1.1 MB · mesh 5.6 MB (from the FaceX README)
- ~17 MB of encrypted weights for the full stack
Figures come from project docs/README; no other benchmarks are claimed.