ArcFace

Proposed by Jiankang Deng, Jia Guo and colleagues at CVPR 2019, ArcFace trains face-recognition networks by adding a fixed angular margin between each embedding and its class center, clustering each person's images on a hypersphere while pushing others apart. The maintained implementation is InsightFace's arcface_torch folder: a config-driven PyTorch trainer for one GPU, an eight-GPU node or multi-node torchrun jobs, with mixed precision, gradient checkpointing, optional NVIDIA DALI loading, ViT backbones, and Partial FC, which samples a fraction of class centers per step so millions of identities still fit in memory. Configs target MS1MV3, Glint360K and WebFace42M, and a converter exports ONNX models. For inference only, the insightface pip package wraps detection, alignment and recognition, and the parent repository stays active. Serious training needs several GPUs. The code is MIT and free for commercial work, but the training data and pretrained models are limited to non-commercial research, and commercial model use needs a separate InsightFace license. Best for teams training their own recognition models.