CADENCE weights

Weights for CADENCE: Temporally Coherent Whole-Body Human Mesh Recovery that Preserves Expressive Hand Motion (Peart, Chauhan, Bodala, Vishwakarma; University of Southampton).

File Stage Size
cadence_per_frame.pt Per-frame whole-body SMPL-X estimator (decoder + projections, and the rank-32 factors of the final-stage C-RADIOv4 adaptation) 260 MB
hand_pathway.pt High-resolution hand pathway and its hand-only backbone LoRA 74 MB
temporal_refiner.pt Kinematic-aware, content-adaptive multi-rate temporal refiner 7.5 MB

The public C-RADIOv4-SO400M backbone is downloaded from NVIDIA via torch.hub at load time; it is not redistributed here.

Usage

git clone https://github.com/keniel123/cadence.git && cd cadence
pip install -r requirements.txt
python scripts/download_weights.py          # fetches these files into weights/
python demo/run_video.py --input clip.mp4 --output out/

SMPL-X must be downloaded separately from https://smpl-x.is.tue.mpg.de.

License

For non-commercial research use only. The weights were trained on datasets (including AGORA, BEDLAM and UBody) and with the SMPL-X body model, whose licences restrict commercial use. The source code is MIT-licensed; see the GitHub repository.

Citation

@article{peart2026cadence,
  title   = {{CADENCE}: Temporally Coherent Whole-Body Human Mesh Recovery
             that Preserves Expressive Hand Motion},
  author  = {Peart, Keniel and Chauhan, Jagmohan and Bodala, Indu P. and Vishwakarma, Shelly},
  journal = {arXiv preprint},
  year    = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support