# OmniHuMo preview samples

The motion observations and captions in `omnihumo-samples.js` are adapted from
[OmniHuMo by L-yiheng and the AnyMo authors](https://huggingface.co/datasets/L-yiheng/OmniHuMo),
released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).
The adapted sample data is distributed under the same license.

These are dataset observations, not motion generated by AnyMo. The viewer uses
recorded axis-angle joint rotations and root translations with a canonical,
fixed-proportion 22-joint skeleton. It does not use the subject's SMPL shape
parameters or reconstruct an SMPL body mesh. Bone proportions are approximate.
The preview does not contain source video, audio, or a licensed SMPL model.

Preview processing uses forward kinematics, frame subsampling, rounding, and a
constant translation to place each clip near the stage origin. The exact source
archive, member paths, sample metadata, and transformations are recorded in the
bundle's `provenance` and sample records. Sixteen complete records near the start
of the archive were selected for varied annotated actions; they are not a
representative sample of the dataset's action distribution. Objects, other
participants, animals, and audio mentioned in captions are not rendered.

Source paper: [AnyMo: Scaling Any-Modality Conditional Motion Generation with
Masked Modeling](https://arxiv.org/abs/2605.29488), Yiheng Li, Zhuo Li, Ruibing Hou,
Yingjie Chen, Hong Chang, Hao Liu, and Shiguang Shan (2026).