qalarc.com / projects / face-tracking-system
Face Tracking System — Unified Platform
One unified face-AI platform: browser-first tracking and recognition, 14 self-trained demographic models, liveness and anti-spoofing, a Python analysis backend, and live avatar generation — fully local, and available for licensing.
Everything Qalarc has built around the human face since January 2026, consolidated into one unified, licensable platform. Four tiers: face-engine, the shared browser-first core (head pose, iris gaze, blink, quality, liveness scoring, 128-d face recognition, rPPG heart-rate) running on MediaPipe FaceMesh + onnxruntime-web with zero install; the Folkus model suite — 14 self-trained ONNX demographic models (MobileNetV2 trained on FairFace+UTKFace on AMD ROCm: gender 92.1% accuracy, age 5.37-year MAE, ethnicity 72.2%, plus ageband, fused age and more); a new Python backend built on uniface (SCRFD detection, FaceMesh478 landmarks, MobileGaze, head pose, BYTETracker persistent IDs) behind FastAPI; and live avatar generation on the local AI media system.
The platform is proven in production. It powers volkus.net's VKVerify age verification — demographics, liveness and MiniFASNet anti-spoofing only, with face recognition hard-disabled by privacy policy — and the smart glasses face tracker demo. You can try the model stack directly in the Volkus Scan demo.
The newest tier works end-to-end: live avatars. Wan 2.2 Animate-14B swaps a driving video's performer for any character (replacement and full-scene animation modes), and Wan 2.2 S2V-14B drives lip-synced talking avatars from speech audio alone — all fully local on an AMD GPU, with a character-canon system keeping identity consistent. The endgame: the tracker's landmark stream becomes the driving signal, so your motion and voice pilot any face.
The platform is available for licensing and partnerships — external teams are already building on it. Contact [email protected] to talk licensing.
Capabilities
Browser face engine: head pose, iris gaze, blink, quality + 30s liveness scoring (MediaPipe FaceMesh, zero install)
14 self-trained Folkus ONNX models — gender 92.1% acc, age 5.37yr MAE, ethnicity 72.2%, ageband + fused age (MobileNetV2, FairFace+UTKFace)
Anti-spoofing: MiniFASNet v2 (1.7MB ONNX) catches photo/screen attacks in-browser
Face recognition with persistent identity — 128-d descriptors, enroll/match/merge (glasses-only privacy policy)
rPPG heart-rate + HRV from face video (green-channel PPG → FFT)
Python backend on uniface: SCRFD detection, FaceMesh478, MobileGaze, BYTETracker persistent IDs
Live face-swap avatars: Wan 2.2 Animate-14B character replacement (replacement + animation modes), fully local
Audio-driven talking avatars: Wan 2.2 S2V-14B lip-sync from speech audio
Tags
Status: active · First built: 2026-08-27 · Last updated: 2026-08-27