UniFace Demo

Interactive demo for UniFace — a comprehensive library for face analysis built on ONNX Runtime.

Features

• Face Detection — RetinaFace, SCRFD, YOLOv5-Face, YOLOv8-Face
• Face Verification — ArcFace, AdaFace, EdgeFace, MobileFace, SphereFace
• Face Analysis — Age, gender, and race prediction (AgeGender, FairFace)
• Landmarks — 106-point (2d106det), 98-point and 68-point (PIPNet)
• Face Mesh — Dense 468-point mesh, or 478-point with irises
• Face Parsing — BiSeNet (19 classes) and XSeg (face mask)
• Emotion — Expression recognition (AffectNet-7 / AffectNet-8)
• Face States — Eyes open, eyeglasses, sunglasses, and mask detection
• Face Quality — Recognition-suitability scoring with eDifFIQA
• Gaze Estimation — Pitch/yaw gaze direction (MobileGaze backbones)
• Head Pose — Pitch/yaw/roll with 3D cube or axis visualization
• Portrait Matting — Trimap-free MODNet alpha matte and background swap
• Face Tracking — ByteTrack multi-face tracking on video input
• Anti-Spoofing — Liveness detection (real vs. fake)
• Face Anonymization — Privacy-preserving blur methods

Running locally

pip install -r requirements.txt
python app.py

Model weights download on first use and are cached afterwards, so the first run of each tab is slower than the rest.

Notes

• torch is required only by the Emotion tab, which loads a TorchScript model. Every other tab runs on ONNX Runtime alone.
• Example images live in assets/, grouped by tab.