Open FAH models estimate hands in real time

Researcher Jinkun Cao released the open FAH (Fast and Accurate Hand Estimation) model family, which estimates hands from a single frame or video at speeds up to 101 fps on a consumer GPU.

Compiled by Sergey KostenchukPublished 2026-11-09Updated 2026-11-09

2026-11-09 Research HuggingFace

🤖 Open FAH models estimate hands in real time

Jinkun Cao released FAH (Fast and Accurate Hand Estimation) — a family of open models for hand estimation from a frame or video. The code is on GitHub (noahcao/FAH) under Apache 2.0, with weights available for research only. The model predicts 2D joints, dense 3D surface points, and object contact, and can be exported to TensorRT and Core ML.

🌍 Offline estimators HaMeR and WiLoR operate at ~4.6-4.7 fps, while FAH-L achieves 4.31 mm PA-MPJPE on HOT3D at 65 fps, and the compact FAH-S reaches 101 fps. This is a ready-made backend for gesture capture, real-to-sim in MuJoCo, and robotics pipelines.

👤 You can clone the repository, download weights from Hugging Face (noahcao/FAH), and run your own photos or videos through the demo. Weights are for research only; registration at mano.is.tue.mpg.de is required for MANO.

Source 1: http://jinkuncao.com/FAH/ Source 2: https://arxiv.org/abs/2611.07711

Sources