FaceX-Zoo is an open-source PyTorch toolbox for face recognition with just under 2,000 stars on GitHub, collecting the field’s important training losses, backbones and evaluation benchmarks into one research codebase from JD’s AI computer-vision team.
Face recognition progress has been driven as much by loss functions as by architectures: margin-based softmax variants taught networks to separate identities more cleanly. FaceX-Zoo gathers those contributions — Adam-Softmax, CircleLoss, MagFace, MV-Softmax, semi-siamese training and more — into a single toolbox where they can be trained and compared under shared conditions, and it includes support for masked face recognition, the problem that suddenly mattered worldwide in 2020. Standard benchmarks such as LFW and MegaFace provide the scoreboard. The accompanying paper appeared in 2021, and the repository’s last updates landed in 2024, so it is best used as a comprehensive reference implementation of the margin-loss era of face recognition: excellent for reproducing results, teaching the techniques, or baselining a new idea, rather than as a maintained SDK. (This listing links to the code repository; the project is also described in its arXiv paper.)
FaceX-Zoo is free and open source. It runs in your own PyTorch environment; the only costs are your compute and any face datasets you are licensed to use for training and evaluation.
Computer-vision researchers and students working on face recognition who want the era’s key loss functions and benchmarks in one place — for reproduction, comparison or teaching — rather than a production face-recognition service.