facex-engine/facex

FaceX is a face-processing library written in C that runs entirely inside a web browser through WebAssembly. The README explains that the full pipeline (face detection, 98-point and 576-point landmarks, face recognition, and a passive anti-spoofing check) ships as roughly 17 MB of encrypted model weights and runs without any server, Python, or GPU. A live demo is hosted on GitHub Pages so visitors can press Start camera in a Chromium browser and try it. Most of the model components were trained by the project's authors from scratch. The README lists each piece with its size and training source: a YuNet-style face detector at 401 KB trained on WIDER FACE, a 98-point landmark model at 1.1 MB trained on WFLW, a 576-point 3D mesh distilled from MediaPipe at 5.6 MB, and recognition models from 0.8 MB to 8.4 MB trained with ArcFace on MS1M, reaching 99.07% on the LFW benchmark. The anti-spoof component is borrowed from MinivisionAI under Apache 2.0. The weights are shipped as AES-256-GCM ciphertext and decrypted in the browser using WebCrypto. The README is honest that this is not real DRM, since a determined attacker can read the decrypted bytes from memory, but it gives friction against casual scraping and lets a SaaS deployment revoke keys per customer. FaceX is positioned as one part of a larger pure-C surveillance stack written by the same team, which also includes a 184 KB H.264 and H.265 decoder called NexusDecode, an HEVC encoder, a surveillance-tuned video format called NXV, and a small neural network inference engine called nn2. The README reports that one Intel i5 CPU can decode 30 RTSP streams and run detection on all of them, supporting around 70 IP cameras on a single core with motion gating. The README walks through what users can build with FaceX: identity verification, face login, access control on edge hardware, proctoring, smart cameras, banking onboarding, and in-store kiosks. A cost comparison table argues that cloud face APIs from AWS, Azure, and Google would cost several thousand dollars per month for a 100,000-user app, while FaceX runs in the user's browser at no per-call cost and avoids sending biometric frames off-device. Small C and shell snippets show how to call the native library and how to serve the browser demo locally. The licence is Apache 2.0.