FAH model weights

Weights for FAH: Fast and Accurate Hand Estimation in Pixels, Interaction and World Space. Code, installation and usage: github.com/noahcao/FAH.

Download

hf download noahcao/FAH --exclude "qc/*" --local-dir ~/.cache/fah/weights

FAH reads the weights from ~/.cache/fah/weights, $FAH_WEIGHTS or --weights DIR. MANO and the SAM 3D Body config are not included; the repository README explains how to get them.

Contents

Path	Contents
fah-s/, fah-b/, fah/ (FAH-L), fah-h/	model.pt (pose network), config.json, contact.pt (contact head), contact_video.pt (video contact model)
shared/detector.pt	hand detector (boxes, handedness and instance masks)
shared/mesh_fitter.pt, shared/mhr_fitter.pt	MANO and MHR-Hand fitters
shared/video_refiner.pt, shared/camera_motion.pt	causal video refiner and camera-motion model
qc/	label-QC lists used by the training recipe (python -m train.fetch_qc, see docs/training.md); not needed for inference
SHA256SUMS	checksums of model checkpoints, configs and QC files (verification below)

Verify downloads

The inference-only command above omits qc/. To verify that download on Linux:

cd ~/.cache/fah/weights
grep -v 'qc/' SHA256SUMS | sha256sum -c -

On macOS, replace sha256sum -c - with shasum -a 256 -c -. For the complete release, download without --exclude "qc/*" and run sha256sum -c SHA256SUMS instead.

Training provenance

The GitHub training guide documents the reproducible entry points and their limits. Historical stage-1 teacher initialisation and exclusion of the auxiliary EgoDex camera-diagnostic episodes are unverified from the files in this release. The corrected code supports explicit fine-tuned-teacher loading, reserves the EgoDex diagnostic candidate pool in both cache building and training, and records new camera training inventories. These safeguards do not change the uploaded checkpoints or establish equivalence of a new training run to them.

Licence

Research and non-commercial use only, under LICENSE. The pose models, fitters, video and contact models derive from SAM 3D Body and DINOv3 and carry the SAM License and the DINOv3 License (reproduced in LICENSE). Built with DINOv3. The detector's provenance and terms are in Part 4 of LICENSE. No dataset images or the MANO model are included.

Citation

@misc{fah2026,
  title  = {Fast and Accurate Hand Estimation in Pixels, Interaction and World Space},
  author = {Cao, Jinkun},
  year   = {2026},
  note   = {Project page and paper link to be added}
}