InfiniteDance

InfiniteDance — Data & Pretrained Weights

This repository hosts the data and weights for InfiniteDance, Scalable 3D Dance Generation Towards in-the-wild Generalization.

Non-commercial use only. Project-owned code, weights, annotations, and data resources are licensed solely for non-commercial academic research, education, and evaluation. Commercial use requires prior written permission. Third-party components and datasets retain their original terms.

The layout mirrors the GitHub repository, so files can be downloaded directly on top of a code checkout.

Contents

All_LargeDanceAR/
├── models/
│   ├── checkpoints/dance_vqvae.pth
│   └── retrievalnet/retrievalnet_audio55_motion264.ckpt
└── output/exp_m2d_infinitedance/best_model_stage2.pt

InfiniteDanceData/
├── DanceVQVAE/body_models/smpl/
├── dance/alldata_new_joint_vecs264/meta/{Mean,Std}.npy
├── partition/
├── styles/all_style_map.json
├── Infinite_MotionTokens_512x1024_3layer_cleandata.tar.gz
├── alldata_new_joint_vecs264_ft_balanced.tar.gz
├── muq_features_test_infinitedance.tar.gz
├── musicfeature_55_allmusic_pure.tar.gz
├── retrieval_s192_l384_style.tar.gz
├── dance/retrievalnet_motion_embeddings.npz
├── dance/evaluation_features_train8235.npz
├── test_eval861_joint_vecs264.tar
└── infinitedance_smplx_smooth.tar

The release includes the 3-layer RVQ motion tokens, cleaned 264-d training features, pretrained InfiniteDance/VQ/RetrievalNet weights, cached retrieval for the test set, live-retrieval corpus embeddings, metric GT features, canonical 861 test GT vectors, and 9,748 project-owned SMPL-X fits.

Each SMPL-X .pkl is a joblib-serialized dictionary with body_pose, global_orient, and transl.

Quick start

git clone https://github.com/MotrixLab/InfiniteDance.git
cd InfiniteDance

pip install -U "huggingface_hub[cli]"
huggingface-cli download huuuuuuuuu/InfiniteDance \
  --repo-type model --local-dir . --local-dir-use-symlinks False

cd InfiniteDanceData
mkdir -p dance music/muq_features
tar -xzf Infinite_MotionTokens_512x1024_3layer_cleandata.tar.gz -C dance/
tar -xzf alldata_new_joint_vecs264_ft_balanced.tar.gz  -C dance/
tar -xzf retrieval_s192_l384_style.tar.gz              -C dance/
tar -xf test_eval861_joint_vecs264.tar                 -C dance/
tar -xf infinitedance_smplx_smooth.tar                 -C dance/
tar -xzf musicfeature_55_allmusic_pure.tar.gz          -C music/
tar -xzf muq_features_test_infinitedance.tar.gz        -C music/muq_features/

cd ../All_LargeDanceAR
bash infer.sh

The main checkpoint contains the complete Llama backbone for inference. Fresh training does not silently random-initialize it: obtain the official Llama-3.2-1B weights from Meta and pass their local path as LLAMA_DIR.

For arbitrary audio, use the GitHub release's utils/extract_muq.py and RetrievalNet/retrieve.py; the latter includes the matching 30-fps Bailando 55-d feature extraction pipeline.

Citation

@misc{li2026infinitedancescalable3ddance,
  title={InfiniteDance: Scalable 3D Dance Generation Towards in-the-wild Generalization},
  author={Ronghui Li and Zhongyuan Hu and Li Siyao and Youliang Zhang and Haozhe Xie and Mingyuan Zhang and Jie Guo and Xiu Li and Ziwei Liu},
  year={2026},
  eprint={2603.13375},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.13375},
}