🎵 MACE-Dance

Motion-Appearance Cascaded Experts for Music-Driven Dance Video Generation

Kaixing Yang¹ · Jiashu Zhu^2,* · Xulong Tang⁵ · Ziqiao Peng¹ · Xiangyue Zhang⁴
Puwei Wang^1,† · Jiahong Wu^2,† · Xiangxiang Chu² · Hongyan Liu^3,† · Jun He<sup>1

1</sup>Renmin University of China   ²AMap, Alibaba   ³Tsinghua University   ⁴Wuhan University   ⁵Malou Tech Inc

^*Project Leader   ^†Corresponding Authors

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

---

✨ Overview

MACE-Dance is the official PyTorch implementation of the SIGGRAPH 2026 paper:

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

Music-driven dance video generation is challenging because it requires simultaneously modeling:

• Motion quality: kinematically plausible and artistically expressive dance motion
• Appearance quality: high-fidelity visual synthesis with strong spatiotemporal consistency

To address this, MACE-Dance decomposes the task into two cascaded experts:

• 🕺 Motion Expert: generates music-aligned 3D dance motion
• 🎨 Appearance Expert: synthesizes the final dance video conditioned on motion and reference appearance

Instead of using 2D keypoints as the intermediate representation, MACE-Dance adopts 3D SMPL motion, which provides better spatial fidelity, cleaner supervision, and stronger robustness for downstream video synthesis.

---

🧩 Repository Structure

MACE-Dance/
├── Expert-Motion/           # Motion Expert: music-to-3D dance motion
├── Expert-Appearance/       # Appearance Expert: motion-guided video synthesis
├── Evaluation-Motion/       # Motion-dimension evaluation
├── Evaluation-Appearance/   # Appearance-dimension evaluation
├── teaser.png
└── README.md

---

📚 MA-Data Dataset

We provide MA-Data, a large-scale dataset for music-driven dance video generation, containing ~70K video clips spanning 116 hours across 20+ dance genres. Please refer to the dataset page for more details.

---

🏋️ Model Weights

The source code for MACE-Dance is fully open-source. For the model weights of the Appearance Expert, please visit the link below to request access or download:

👉 Click here to access MACE-Dance Model Weights

---

📏 Evaluation Protocol

We provide a motion–appearance evaluation protocol for music-driven dance video generation, including motion quality assessment based on ViTPose keypoints and appearance quality assessment based on VBench. Please refer to Evaluation-Motion and Evaluation-Appearance for details.

---

📌 Notes

• The repository is organized into expert modules and evaluation modules.
• Please check the subfolder READMEs for environment setup, inference, and evaluation details.
• Some released example files are for demonstration only; please replace them with your own predictions / ground-truth files during evaluation.

---

📄 Citation

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

@inproceedings{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 Wu, Jiahong and Chu, Xiangxiang and Liu, Hongyan and He, Jun},
  booktitle={Proceedings of the ACM SIGGRAPH Conference},
  year={2026}
}

---

🙏 Acknowledgement

This work was supported in part by the National Nature Science Foundation of China under Grants 62436010, 72572090, 62572474, and 62172421, and in part by the Tsinghua University School of Economics and Management Research Grant.