Tevior/TopX-OOD

TopX

TopX generates motion for an unseen skeleton from a text prompt and the skeleton file alone: its kinematic tree, rest pose and joint names. No clips of the character, templates, retargeting or fine-tuning are needed. It is a 73M-parameter flow-matching transformer with attention factorised over joints and frames; across joints, queries and keys are rotated by HK-RoPE, a heat-kernel rotary encoding of the skeleton graph. One model animates skeletons from 4 to 117 joints and generates up to 240 frames at 30 fps.

Code: https://github.com/CHDTevior/TopX-OOD-Code · Project page: https://chdtevior.github.io/TopX-page/ · Paper: arXiv:2610.34321

Files

file	content	sha256
topx.pt	{"config": architecture and normalisation, "model": state dict} (293 MB, float32)	f6d8e086235d1f99069cce47a2a7e5d59141a9442b7b5b2880937716b0ee405a

Usage

git clone https://github.com/CHDTevior/TopX-OOD-Code.git && cd TopX-OOD-Code
pip install -r requirements.txt
huggingface-cli download Tevior/TopX-OOD topx.pt --local-dir weights
bash demo/run_demos.sh                                   # nine demo skeletons, no text encoder needed
python scripts/generate.py --ckpt weights/topx.pt --skeleton my_rig.bvh --up +Y --forward +Z \
    --text "An object walks forward and sits down." --frames 120 --out out/

The output is a BVH file of the input hierarchy (in its own axes and units), an .npz with joint positions and rotations, and a GIF preview. Prompts and joint descriptions are encoded with LLM2Vec (McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised, built on the gated meta-llama/Meta-Llama-3-8B-Instruct).

Training data

The training split of the SkelShift benchmark: the Objaverse-XL subset of UniML3D, the AniMo4D animal library with HumanML3D, further animations and species from the same game library as AniMo4D, and the Truebones zoo (139,102 clips on 5,127 skeletons with 6 to 142 joints). See the paper for the split and the evaluation.

License

The weights are released under CC BY-NC 4.0: non-commercial use with attribution. The code is MIT-licensed. Licences of the training data and of the text encoder apply to their respective parts.

Citation

@article{sun2026topx,
  title   = {TopX: Text-to-Motion for Unseen Skeletons from the Skeleton File Alone},
  author  = {Sun, Tengjiao and Fang, Pengcheng and Zhan, Xiaoyu and Fu, Dongjie and Cai, Xiaohao and Kim, Hansung},
  journal = {arXiv preprint arXiv:2610.34321},
  year    = {2026}
}