facetorch

Why we included this project

Facetorch bundles community-built face analysis models into a single PyTorch toolkit, covering detection, landmarks, expression, verification, action units, and 3D alignment. Because the models are serialized with torch.export into .pt2 artifacts, you can run inference without the original model source code, with dynamic batch sizes and torch.compile support. Hydra and OmegaConf handle configuration, and the project ships Docker images and a Colab notebook, so it is easy to test on CPU or GPU before committing to a deployment. Teams building identity checks, emotion analytics, or deepfake screening will find a curated model set that saves them from wiring together several separate repositories, though the Linux x86-64 focus and pinned PyTorch versions mean you should check your target environment first.