---
license: cc-by-4.0
pipeline_tag: image-to-video
---

# MACE-Dance: Motion-Appearance Cascaded Experts for Music-Driven Dance Video Generation

[**Paper**](https://arxiv.org/abs/2512.18181) | [**Project Page**](https://macedance.github.io/) | [**GitHub**](https://github.com/AMAP-ML/MACE-Dance)

MACE-Dance is a cascaded expert framework for music-driven dance video generation. It explicitly decouples motion generation and appearance synthesis to produce kinematically plausible, artistically expressive, and visually coherent dance videos.

The framework consists of two main components:
- **Motion Expert**: Performs music-to-3D motion generation (SMPL) using a diffusion model with a BiMamba-Transformer architecture.
- **Appearance Expert**: Synthesizes the final dance video conditioned on the motion and a reference image, ensuring visual identity and spatiotemporal coherence.

## 📦 Pretrained Weights

This repository stores the pretrained weights used in **MACE-Dance**, including both **expert models** and **evaluation models**.

### 🗂️ Directory Structure

```bash
weight/
├── Evaluation-Appearance/
│   └── Evaluation-Appearance.7z
├── Evaluation-Motion/
│   └── Evaluation-Motion.7z
├── Expert-Appearance/
│   └── Expert-Apprearance.7z
└── Expert-Motion/
    └── Expert-Motion.7z
```

### 📘 Description

- 🎭 **Expert-Appearance**: Pretrained weights for the **Appearance Expert** (motion-guided video synthesis).
- 🕺 **Expert-Motion**: Pretrained weights for the **Motion Expert** (music-to-3D dance motion).
- 📊 **Evaluation-Appearance**: Weights used for **appearance-related evaluation** (based on VBench).
- 📈 **Evaluation-Motion**: Weights used for **motion-related evaluation** (based on ViTPose).

## 📝 Notes

- Each subfolder contains a compressed **`.7z`** package. Please extract the corresponding file before use.
- Make sure the extracted weights are placed in the expected paths for training, inference, or evaluation as specified in the [official code repository](https://github.com/AMAP-ML/MACE-Dance).
- Use **Expert-*** weights for model inference and **Evaluation-*** weights for metric computation pipelines.

## 📄 Citation

If you find this project useful, please consider citing the paper:

```bibtex
@article{yang2026macedance,
  title={MACE-Dance: Motion-Appearance Cascaded Experts for Music-Driven Dance Video Generation},
  author={Yang, Kaixing and Zhu, Jiashu and Tang, Xulong and Peng, Ziqiao and Zhang, Xiangyue and Wang, Puwei and Jiahong Wu and Chu, Xiangxiang and Liu, Hongyan and He, Jun},
  journal={ACM Transactions on Graphics (SIGGRAPH 2026)},
  year={2026}
}
```