RigMo: Unifying Rig and Motion Learning for Generative Animation
Paper • 2601.06378 • Published • 12
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is a preprocessed derivative of DeformingThings4D, Objaverse-XL, and TrueBones, provided for non-commercial academic research only. By requesting access you agree to (1) use the data solely for non-commercial research, (2) comply with the original licenses and terms of DeformingThings4D, Objaverse-XL, and TrueBones, and (3) cite the RigMo paper and the original datasets in any resulting work. The authors make no warranty and accept no liability for use of this data.
Log in or Sign Up to review the conditions and access this dataset content.
RigMo Mesh-Motion Dataset
Preprocessed mesh-sequence data used to train the RigMo-VAE from RigMo: Unifying Rig and Motion Learning for Generative Animation.
- 📄 Paper: https://arxiv.org/abs/2601.06378
- 🌐 Project page: https://rigmo-page.github.io/
- 💻 Code: https://github.com/haoz19/RigMo
Scale: ~18,985 sequences · ~534k .npz frames · ~46 GiB.
Download
The data ships as 10 .tar.zst archives (one per group) so the half-million
small frame files transfer efficiently. Download them, extract into a single
folder, and point the training config at that folder.
# 1. Download all archives (requires `pip install huggingface_hub` and access approval)
huggingface-cli download haoz19/RigMo-data \
--repo-type dataset --local-dir rigmo_data_archives
# 2. Extract every archive into ./rigmo_data (needs `zstd` + `tar`)
mkdir -p rigmo_data
for f in rigmo_data_archives/*.tar.zst; do
tar -I zstd -xf "$f" -C rigmo_data
done
# 3. (optional) remove the archives to reclaim space
# rm -rf rigmo_data_archives
After extraction you get the layout the training code expects:
rigmo_data/
├── deformingthings4d/ # sequences derived from DeformingThings4D
├── objxl_rendered_0_2500/ # Objaverse-XL render shards (8 dirs)
├── objxl_rendered_2500_5000/
├── ...
└── val/ # held-out validation split (100 sequences)
Then train (see the code repo):
python train.py --config configs/rigmo_vae_temporal_single_node.yaml --train \
data.root_dir=/abs/path/to/rigmo_data
| Archive | Description |
|---|---|
deformingthings4d.tar.zst |
Sequences derived from DeformingThings4D |
objxl_rendered_*.tar.zst |
Sequences derived from Objaverse-XL renders (8 shards) |
val.tar.zst |
Held-out validation split (100 sequences) |
Format
Each sequence is a directory of per-frame .npz files:
<sequence_name>/
├── frame_0000.npz # vertices [N, 3] float32 · neighbor_idx [N, k] int64
├── frame_0001.npz
└── ...
| Key | Shape | Description |
|---|---|---|
vertices |
[N, 3] float32 |
per-frame vertex positions (here N = 5000) |
neighbor_idx |
[N, k] int64 |
per-vertex mesh neighbors (mesh topology) |
Sequences are normalized at load time so the first frame's bounding box maps to a
unit cube centered at the origin. See the
training code (FullMeshMotionNPZ-datamodule)
for exact loading details. The training data module recursively discovers sequence
directories and reserves val/ (and test/, if present) as held-out splits.
Licensing & attribution
This is a derivative dataset for non-commercial academic research only. It is built from:
- DeformingThings4D — academic / non-commercial; subject to its original terms.
- Objaverse-XL — ODC-BY; individual assets retain their own licenses.
- TrueBones — subject to TrueBones' own terms.
You must comply with all original dataset licenses. Access is gated; requesting access constitutes agreement to the terms above.
Citation
@article{zhang2026rigmo,
title = {RigMo: Unifying Rig and Motion Learning for Generative Animation},
author = {Zhang, Hao and Luo, Jiahao and Wan, Bohui and Zhao, Yizhou and Li, Zongrui
and Vasilkovsky, Michael and Wang, Chaoyang and Wang, Jian and Ahuja, Narendra
and Zhou, Bing},
journal = {arXiv preprint arXiv:2601.06378},
year = {2026}
}
- Downloads last month
- 2