TunaDance

TunaDance

Music-to-dance generation with a Gradio web UI and single-command CLI.

TunaDance builds on FineDance (ICCV 2023), a diffusion-based model that generates full-body 3D dance from music. This fork finetunes the original model on additional data and for more epochs beyond the original 2000, and adds a user-friendly interface layer and macOS support so you can go from an audio file to a rendered dance video without touching the inference internals.

[Original Paper] | [Original Project Page] | [Original Repo]

<img src="teaser/teaser.png">

What's New (vs. upstream FineDance)

• Gradio Web UI (app.py) — Upload music in the browser, get a dance video back. No CLI knowledge required.
• Single-command CLI (generate_dance.py) — One command handles the full pipeline: audio feature extraction, diffusion sampling, SMPLX rendering, and audio-video muxing.
• macOS / MPS support — Updated render.py, vis.py, and inference code to run on Apple Silicon via MPS, with a dedicated environment_macos.yaml.
• Accepts any audio format — Automatically converts input to WAV via ffmpeg (.mp3, .wav, .m4a, .flac, .ogg, etc.).
• Finetuned checkpoint — Finetuned the original FineDance model on additional data and for more epochs beyond the original 2000, improving dance quality and diversity.
• Cleaned-up repo — Removed wandb logs, debug scripts, and hardcoded paths.

Model Details

Quick Start

Prerequisites

# Install conda environment
conda env create -f environment.yaml        # Linux/CUDA
conda env create -f environment_macos.yaml   # macOS (Apple Silicon)

conda activate FineNet

Download the pretrained checkpoint and SMPLX model from Google Drive and place them under assets/.

Web UI (Recommended)

python app.py

Open http://127.0.0.1:7861 in your browser. Upload a music file and click Generate Dance.

Command Line

python generate_dance.py /path/to/music.mp3

Output is saved to output/<songname>_dance.mp4. Use --output for a custom path:

python generate_dance.py /path/to/music.mp3 --output my_dance.mp4

Output Specs

How It Works

1. Audio conversion — Converts input to WAV if needed via ffmpeg
2. Feature extraction — Slices audio into 4s windows (2s stride), extracts 35-dim features using librosa
3. Dance generation — Diffusion model generates SMPLX motion sequence from audio features
4. Rendering — Converts motion to SMPLX meshes, renders 900 frames at 30fps with pyrender
5. Muxing — Merges rendered video with original audio via ffmpeg

Training

Only needed to train from scratch. The pretrained checkpoint is included.

python data/code/pre_motion.py                              # preprocess
accelerate launch train_seq.py --batch_size 32 --epochs 200 # train

Key flags:

• --batch_size — Default 400; reduce to 32 or lower for Mac MPS
• --epochs — Default 2000
• --checkpoint — Resume from a saved checkpoint

FineDance Dataset

The dataset (7.7 hours) is available from Google Drive or Baidu Cloud. Place it under ./data.

import numpy as np
data = np.load("motion/001.npy")
smpl_poses = data[:, 3:]   # joint rotations
smpl_trans = data[:, :3]   # root translation

Two dataset splits are provided:

• FineDance@Genre (recommended) — Broader genre coverage in the test set
• FineDance@Dancer — Splits by dancer identity

Project Structure

TunaDance/
├── app.py                   # Gradio web UI  [NEW]
├── generate_dance.py        # End-to-end CLI [NEW]
├── environment_macos.yaml   # macOS conda env [NEW]
├── train_seq.py             # Training script
├── test.py                  # Original inference script
├── render.py                # SMPLX mesh rendering (updated for MPS)
├── vis.py                   # Skeleton/FK utilities (updated for MPS)
├── args.py                  # CLI argument definitions
├── assets/
│   ├── checkpoints/
│   │   └── train-2000.pt    # Pretrained model (2000 epochs)
│   └── smpl_model/
│       └── smplx/
│           └── SMPLX_NEUTRAL.npz
├── model/
│   ├── model.py             # SeqModel (transformer decoder)
│   └── diffusion.py         # Gaussian diffusion
├── dataset/
│   └── FineDance_dataset.py
└── data/
    └── finedance/           # Training data (music + motion pairs)

Acknowledgments

This project is built on FineDance by Li et al. We thank the original authors for their work.

Upstream acknowledgments: EDGE, MDM, Adan, Diffusion, SMPLX.

Citation

@inproceedings{li2023finedance,
  title={FineDance: A Fine-grained Choreography Dataset for 3D Full Body Dance Generation},
  author={Li, Ronghui and Zhao, Junfan and Zhang, Yachao and Su, Mingyang and Ren, Zeping and Zhang, Han and Tang, Yansong and Li, Xiu},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={10234--10243},
  year={2023}
}