InfiniteDance — Data & Pretrained Weights

This repo hosts the data and weights for InfiniteDance (Scalable 3D Dance Generation Towards in-the-wild Generalization).

The directory layout here mirrors the GitHub repo — every file's path on HF is exactly where it should live locally. You can drop this repo on top of a clone of the code repo without renaming anything.

Contents (HF path → local path are identical)

models/checkpoints/dance_vqvae.pth                    # 462 MB → All_LargeDanceAR/models/checkpoints/
output/exp_m2d_infinitedance/best_model_stage2.pt     # 2.3 GB → All_LargeDanceAR/output/exp_m2d_infinitedance/
                                                      #          (full LLaMA-3.2-1B backbone is INSIDE this ckpt;
                                                      #           no download from Meta is required)

InfiniteDanceData/
├── DanceVQVAE/body_models/smpl/                      # 40 MB → SMPL body models for visualization
├── dance/alldata_new_joint_vecs264/meta/
│   ├── Mean.npy                                      # 264-d motion-feature mean
│   └── Std.npy                                       # 264-d motion-feature std
├── partition/                                        # train / val / test splits
├── styles/all_style_map.json                         # genre label map
├── Infinite_MotionTokens_512_vel_processed.tar.gz    #  14 MB → motion tokens (training)
├── muq_features_test_infinitedance.tar.gz            # 2.6 GB → MuQ features (test set)
├── musicfeature_55_allmusic_pure.tar.gz              # 3.0 GB → 55-d music feats (BA metric)
└── retrieval_s192_l384_style.tar.gz                  # 839 MB → top-K retrieval cache

Quick start

# 1. clone the code repo
git clone https://github.com/MotrixLab/InfiniteDance.git
cd InfiniteDance

# 2. download everything from HF directly on top of the clone — paths match,
#    so files land in the right place automatically
pip install -U "huggingface_hub[cli]"
huggingface-cli download huuuuuuuuu/InfiniteDance \
    --repo-type model \
    --local-dir . \
    --local-dir-use-symlinks False

# 3. extract the four tarballs in place
cd InfiniteDanceData
mkdir -p dance music/muq_features
tar -xzf Infinite_MotionTokens_512_vel_processed.tar.gz -C dance/
tar -xzf retrieval_s192_l384_style.tar.gz              -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 ..

# 4. run inference
cd All_LargeDanceAR
bash infer.sh

Expected layout after extraction

InfiniteDance/
├── All_LargeDanceAR/
│   ├── models/checkpoints/dance_vqvae.pth
│   ├── models/Llama3.2-1B/config.json                 # architecture only (shipped with code)
│   └── output/exp_m2d_infinitedance/best_model_stage2.pt
└── InfiniteDanceData/
    ├── DanceVQVAE/body_models/smpl/
    ├── dance/
    │   ├── alldata_new_joint_vecs264/meta/{Mean,Std}.npy
    │   ├── Infinite_MotionTokens_512_vel_processed/
    │   └── retrieval_s192_l384_style/
    ├── music/
    │   ├── muq_features/test_infinitedance/
    │   └── musicfeature_55_allmusic_pure/
    ├── partition/
    └── styles/

Notes

• LLaMA weights are bundled in the released checkpoint. You do not need to download anything from Meta. The code repo ships only the architecture config.json so the model can be re-instantiated; the weights are loaded from best_model_stage2.pt.
• The pretrained model uses a per-frame MLP dance bridge and an n_bins=2 cadence MoE music bridge, with anti-collapse strict decoding enabled by default (see infer.sh).
• Not yet released here: the full 264-dim normalized motion features (alldata_new_joint_vecs264/, needed to train from scratch) and the SMPL-X GT joints (ourData_smplx_22_smooth_new/, needed to compute FID-k / FID-m / Div-k / Div-m). What you can reproduce today: inference on the released MuQ test set, inference on your own audio, the Beat-Align metric, and retrieval ablations.

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},
}