Home

FaceX Wiki

Full face stack that runs in the browser. Detection, 576-point 3D mesh, recognition, anti-spoof, smile, pulse — all WebAssembly, zero servers, weights are AES-256-GCM encrypted.

🎬 Live Demo · Source · Apache 2.0

---

Pages

Browser Quickstart	30-line copy-paste recipe to embed FaceX in any web app
Training Recipes	Reproduce every shipped model from scratch (recognition, detector, landmarks, mesh, smile)
nn2 Architecture	The pure-C inference engine: GEMM 6×32, AVX-512 ASM, runtime SIMD dispatch
Encrypted Weights	Serve AES-encrypted models without leaking IP, with backend key endpoints
Comparison vs Alternatives	FaceX vs AWS Rekognition / Azure Face / MediaPipe / InsightFace / FaceTec
Roadmap	What ships next: ESP32-P4, iMX NPU, Apple Silicon SME, smile/age/emotion

Stack snapshot

Component	Size	Speed	Source
Face detector (ours, WIDER FACE)	401 KB	~1 ms	this repo
98-pt landmarks (ours, WFLW)	1.1 MB	~4 ms	this repo
576-pt 3D mesh (ours, MP distill)	5.6 MB	~8 ms	this repo
Recognition (ours, 4 variants)	0.8–8.4 MB	1.4–3 ms	this repo
MiniFASNet anti-spoof	2 × 1.7 MB	1.4 ms (ensemble, nn2)	MinivisionAI Apache 2.0
Smile classifier (ours)	187 KB	<2 ms	this repo

All weights AES-256-GCM encrypted, decrypted in-browser via WebCrypto.

Quick facts

• 6 trained classifiers, F1 0.95+ across the binary ones
• 1.5–2× faster than ONNX Runtime via nn2
• 0 servers required for the browser pipeline
• ~$0/mo at any traffic level (vs ~$3K/mo at 100 K MAU on AWS Rekognition)
• Compliance-friendly: frames never leave the device

Contact: bauratynov@gmail.com