InsightFace

Face detection, landmark alignment, recognition embeddings, and 3D reconstruction all live in this long-running research toolbox led by Jia Guo and Jiankang Deng, best known for the ArcFace loss and the RetinaFace and SCRFD detectors. The repository holds PyTorch and MXNet training recipes, IJB and MegaFace evaluation, and the pip-installable insightface package, which runs ONNX models on CPU, CUDA, or CoreML and fetches model packs automatically. Releases during 2026 added a desktop GUI; PrivateFrame, which blurs or pixelates faces in video entirely on your own machine; an optional RGB liveness check; PersonAnalysis with body re-identification; and InsightFace Server, a self-hosted container with a browser dashboard, REST endpoints, and SQLite-backed face search on Linux CPU or NVIDIA GPU. Licensing needs care: the code is MIT, but the published models and training data are limited to non-commercial research, and commercial use of recognition packs such as buffalo_l, the inswapper face swap models, or the InspireFace SDK requires a paid license. Suited to researchers and to teams prepared to license models for production.