Routed and Denoising-Aware Music Conditioning for Dance Video Generation
RoDA-Dance is a music-driven dance video generation framework for improving music-motion coordination, motion diversity, and motion quality while maintaining visual fidelity, subject consistency, and temporal stability. The project focuses on effective modeling and coordination of music conditioning in diffusion-based video generation.
Project page: https://minervaw26.github.io/RoDA-Dance/
Music-driven dance video generation requires models to follow semantic instructions, respond to musical dynamics, and preserve visual fidelity. Existing end-to-end methods are limited by local music modeling, strong audio-motion binding, and imbalanced text-music conditioning across denoising stages.
RoDA-Dance addresses these challenges with:
- Global-Local Music Conditioning for combining local rhythmic cues with segment-level musical context.
- Null-Routed Audio Cross-Attention for controlling audio attention strength while preserving the relative distribution among real audio tokens.
- Denoising-Aware Audio Rebalancing for coordinating text and music guidance across denoising stages.
@inproceedings{wang2026roda,
title={RoDA-Dance: Routed and Denoising-Aware Music Conditioning for Dance Video Generation},
author={Meng Wang, Zichao Nie, Xu He, Liyang Chen, Haiwei Xue, Runnan Li and Zhiyong Wu},
booktitle={ICASSP},
year={2027}
}