--- tags: - music-to-dance - dance-generation - 3d-human-motion - motion-generation - meanflow - bimamba ---

๐ŸŒธ FlowerDance

MeanFlow for Efficient and Refined 3D Dance Generation

Paper Project Page Conference Hugging Face

FlowerDance teaser

> **Abstract**: Music-to-dance generation translates auditory signals into expressive human motion, yet existing approaches still struggle to balance refined 3D motion quality with strict inference budgets. FlowerDance is designed for both physically plausible, artistically expressive motion and efficient generation in speed and memory usage. > > FlowerDance combines MeanFlow with Physical Consistency Constraints for high-quality few-step sampling, and uses a lightweight non-autoregressive BiMamba backbone with Channel-Level Fusion for long-horizon music-to-dance synthesis. It also supports motion editing through time-decayed soft masking, enabling users to refine generated dance sequences interactively.

๐ŸŽ‰ FlowerDance has been accepted to ECCV 2026! ๐ŸŽ‰
โœจ Training and inference code are now available. โœจ

--- ## ๐Ÿš€ Code The complete training, evaluation, and inference code is maintained in the [GitHub repository](https://github.com/XulongT/FlowerDance). ### ๐Ÿ› ๏ธ Set up the Environment To set up the necessary environment for running this project, follow the steps below: 1. **Clone the repository** ```bash git clone https://github.com/XulongT/FlowerDance.git cd FlowerDance ``` 2. **Create a new conda environment** ```bash conda create -n Flower_env python=3.10 conda activate Flower_env ``` 3. **Install PyTorch (CUDA 12.8)** ``` pip install torch==2.7.1+cu128 torchvision==0.22.1+cu128 torchaudio==2.7.1+cu128 \ --index-url https://download.pytorch.org/whl/cu128 ``` 4. **Install remaining dependencies** ```bash pip install -r requirements.txt ``` --- ## ๐Ÿ“ฆ Download Resources - Download the complete **preprocessed data archive** from [Hugging Face](https://huggingface.co/datasets/xlt99/FlowerDance-Preprocessed/resolve/main/data.7z?download=true) and extract it so that the preprocessed files are located under `./data/`. - The **pretrained checkpoint** is hosted in this model repository. Custom-music inference downloads it automatically when `--checkpoint` is omitted. --- ## ๐Ÿงฉ Directory Structure After downloading the necessary files, ensure the directory structure follows the pattern below: ``` FlowerDance/ โ”‚ โ”œโ”€โ”€ data/ โ”œโ”€โ”€ dataset/ โ”œโ”€โ”€ model/ โ”œโ”€โ”€ runs/ โ”œโ”€โ”€ requirements.txt โ”œโ”€โ”€ args.py โ”œโ”€โ”€ EDGE.py โ”œโ”€โ”€ train.py โ”œโ”€โ”€ test.py โ”œโ”€โ”€ inference.py โ”œโ”€โ”€ inpaint.py โ””โ”€โ”€ vis.py ``` --- ## ๐Ÿ‹๏ธ Training ```bash export WANDB_MODE=offline accelerate launch train.py --batch_size 128 --epochs 4000 --feature_type baseline ``` --- ## ๐Ÿ“ Evaluation ### ๐Ÿงช Evaluate the Model To evaluate the model: ```bash python test.py --batch_size 128 ``` --- ## ๐ŸŽต Inference Generate a genre-conditioned dance sequence of a custom length from your own music: ```bash python inference.py path/to/music.wav \ --genre Hiphop \ --duration 32 ``` List all supported dance genres: ```bash python inference.py --list-genres ``` The checkpoint is downloaded automatically from [Hugging Face](https://huggingface.co/xlt99/FlowerDance) when `--checkpoint` is omitted. Generated motions are saved to `inference_outputs/`. --- ## ๐Ÿ™ Acknowledgements This code is standing on the shoulders of giants. We want to thank the following contributors that our code is based on: [EDGE](https://github.com/Stanford-TML/EDGE), [Adan-pytorch](https://github.com/lucidrains/Adan-pytorch), [denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch), [Mamba](https://github.com/state-spaces/mamba), [causal-conv1d](https://github.com/Dao-AILab/causal-conv1d), and [fairmotion](https://github.com/facebookresearch/fairmotion). The preprocessed data builds on [AIST++](https://github.com/google/aistplusplus_api) and [FineDance](https://github.com/li-ronghui/FineDance). --- ## ๐Ÿ“„ Citation ```bibtex @article{yang2025flowerdance, title={FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation}, author={Kaixing Yang and Xulong Tang and Ziqiao Peng and Xiangyue Zhang and Puwei Wang and Jun He and Hongyan Liu}, journal={arXiv preprint arXiv:2511.21029}, year={2025} } ```