kinematic-frequency-motion-weights

Kinematic Frequency Motion — Released Weights

Learning Kinematic Frequency-Aware Disentanglement for Motion Style Transfer and Editing, SIGGRAPH Asia 2026.

These are the five checkpoints used by the paper figures and the live demo. The content motion is encoded by the motion VAE. The style motion is split by the IMF extractor into high-, mid-, and low-frequency bands. High and mid bands supply the fine style. A summary of the low band is encoded by MotionCLIP into the coarse-style token. The dual-style denoiser writes a latent, and the VAE decoder turns it back into a motion.

• Project page
• Code
• Live demo

Files

Use the five files together. About 2.3 GB in total.

SHA256SUMS lists the SHA-256 and byte size of each file. CLIP ViT-B/32 is not in this repository. The clip package downloads it on the first run.

Usage

hf download randong/kinematic-frequency-motion-weights --local-dir checkpoints

The Colab notebook and the live demo load these files with huggingface_hub. A local checkout can point checkpoints/ at the downloaded directory. Inference reads raw HumanML-263 features, shape [T, 263], at 20 Hz, and leaves them unnormalized.

Paper settings for the shipped clips: turning walk with a periodic dance kick uses coarse scale 2.5 and fine scale 1.5 (fine 10 for the dance accent). Stair walk with an energetic style uses coarse 2.5 and fine 3.

Citation

@inproceedings{dong2026kinematic,
  title={Learning Kinematic Frequency-Aware Disentanglement for Motion Style Transfer and Editing},
  author={Dong, Ran and Xie, Haoran and Yang, Xi},
  booktitle={SIGGRAPH Asia 2026 Conference Papers},
  year={2026},
  note={to appear}
}