The animal data here derives from Truebones Zoo. Please get it from Truebones yourself first — https://truebones.gumroad.com/l/skZMC — it is pay-what-you-want, so a free download is perfectly fine.

What matters is not that you paid, but that you obtained it from Truebones directly and are therefore bound by their terms, which prohibit redistributing the motion files. Requests are reviewed manually.

MoCapAnything V2 — Training Data

Training data for MoCapAnything V2: end-to-end motion capture for arbitrary skeletons from monocular video.

💻 Code: github.com/phongdaot/MocapAnything
🏋️ Weights: kehong/MoCapAnythingV2-weights
🎮 Demo data (ungated, small): kehong/MoCapAnythingV2-data-sample
📄 Paper: arXiv:2604.28130

⚠️ Access requires your own Truebones licence

This is a gated dataset. The animal half derives from Truebones Zoo, a commercial motion library that Truebones distributes on a pay-what-you-want basis. Their terms are royalty-free for commercial use but prohibit redistributing the motion files, so access here is granted only to people who have obtained it from Truebones themselves and are bound by those terms. Requests are reviewed manually.

If you only want to run the demo or reproduce inference, you do not need this dataset — use the ungated data-sample instead.

Path	What	Size
`zoo1030/bvh/`	Animal skeletons + motion, 3 camera yaws per sequence	1.9 GB
`zoo1030/bvh_pose/`	Extracted joint positions / rotations (`.npz`), the training targets	3.8 GB
`zoo1030/characters_fix_facezplus/`	Rigged meshes + skinning weights, face-`+z` aligned	174 MB
`zoo1030/ref_images/`, `ref_images_y90/`	Per-species reference frames (front / side)	366 KB
`zoo1030/cache/`	Per-species global-scale cache	5.8 MB
`zoo1030/species_info_dict.npy`	Joint names, topology, per-species metadata	12 MB
`zoo1030/selected_test_split_release.json`	Train/test split (`seen` / `rare` / `unseen`)	3.4 KB
`obj1k/…`	Same layout for the Objaverse-derived object subset	~19 GB

Not included: image features

npz_train_image_only/ (DINOv2 per-frame features, ~1.2 TB across both subsets) is not distributed. Regenerate it from the rigged meshes with the released pipeline:

# render → extract features; see preprocess/README.md for the full chain
python preprocess/render_bvh_videos.py    ...
python preprocess/extract_image_only.py   ...

Everything the training loader needs beyond those features is here.

Usage

hf download kehong/MoCapAnythingV2-data --repo-type dataset --local-dir ./datasets
torchrun --nproc_per_node=8 -m train.video2pose2rot_multidata \
    --config configs/train/train_video2pose2rot_multidata.yaml

The config expects this layout under datasets/, matching the paths in configs/train/train_video2pose2rot_multidata.yaml. See RUN.md → Metrics & Reproduction for the training and evaluation commands and the numbers to expect.

Downloading only part of it

The full download is ~25 GB across ~57k files. Files are stored unpacked, so you can take just the parts you need:

# animals only (~6 GB)
hf download kehong/MoCapAnythingV2-data --repo-type dataset \
    --local-dir ./datasets --include "zoo1030/**"

# skeletons + metadata, without the 8.4 GB of pose targets
hf download kehong/MoCapAnythingV2-data --repo-type dataset \
    --local-dir ./datasets --exclude "*/bvh_pose/**"

# a single species
hf download kehong/MoCapAnythingV2-data --repo-type dataset \
    --local-dir ./datasets --include "zoo1030/bvh/Horse*/**" "zoo1030/bvh_pose/Horse*/**"

Licensing

Animal data (zoo1030/) derives from Truebones Zoo © Truebones Motions Animation Studios. Their terms are royalty-free and permit commercial use, and contain no clause about derivative works or trained models; they prohibit redistributing the motion files. Access here is gated on your own Truebones licence. Credit Truebones as their terms request.

Object data (obj1k/) derives from Objaverse; individual assets retain their own (mostly Creative Commons) licences and attribution requirements.

The released model weights are MIT and carry none of these constraints — see kehong/MoCapAnythingV2-weights.

Citation

@article{gong2026mocapanythingv2,
  title   = {MoCapAnything V2: End-to-End Motion Capture for Arbitrary Skeletons},
  author  = {Gong, Kehong and Wen, Zhengyu and Phong, Dao Thien and
             Xu, Mingxi and He, Weixia and Wang, Qi and Zhang, Ning and
             Li, Zhengyu and Hou, Guanli and Lian, Dongze and He, Xiaoyu and
             Zhang, Mingyuan and Zhang, Hanwang},
  journal = {arXiv preprint arXiv:2604.28130},
  year    = {2026}
}