FlowDance: Music-Driven Dance Video Generation with Parallel Pose and RGB Streams

Genying Li1, 2*, Boda Lin1*, Jiachen Li1, Zijian Jia1, 2, Haojie Zheng2, 3, Yiming Wang1, Shuchen Weng2†, Si Li1†
1School of Artificial Intelligence, Beijing University of Posts and Telecommunications
2Beijing Academy of Artificial Intelligence
3School of Software and Microelectronics, Peking University

*Equal contribution    †Corresponding author
FlowDance pipeline and qualitative comparison results

Abstract

Music-driven dance video synthesis aims to animate a reference person according to a given music clip. The task is challenging because it requires a model to jointly learn music-to-motion correspondence, identity-preserving human animation, temporal coherence, and visually realistic video generation. We present FlowDance, a music-driven dance video generation framework that integrates explicit motion modeling with reference-preserving visual synthesis through parallel pose and RGB streams. We further introduce timestep-aware pose injection to adapt structural guidance across denoising steps and persistent identity injection to preserve the reference appearance over long video. To support this task, we further build a popularity-curated, high-resolution in-the-wild dance video dataset with synchronized music, RGB videos, 3D body motion, camera parameters, and projected 2D pose annotations. Extensive experiments show that FlowDance achieves strong performance in both dance motion generation and music-driven dance video synthesis.

Method

Overview of the FlowDance framework with feature extraction, parallel RGB and pose streams, and M2V aligners

FlowDance models dance motion and visual rendering in parallel within a pretrained video diffusion backbone. The feature-extraction stage provides identity, visual, text, and audio conditioning to two coupled streams: a pose stream that predicts music-aligned 2D motion, and an RGB stream that synthesizes the target person under continuous structural guidance. Intermediate pose features guide the corresponding RGB blocks through timestep-aware M2V Aligners, so the streams remain coupled throughout denoising rather than passing motion through a one-way cascade. The figure also details the Face Encoder and M2V Aligner modules.

Face-token injection and face-masked reconstruction reinforce the appearance of the reference person, while the pose stream avoids an explicit 3D-to-2D projection step. This design separates motion planning from appearance synthesis without losing temporal and structural correspondence, enabling expressive, identity-preserving dance video generation from a reference image and music.

Result

Same Reference, Different Music

Contemporary
Street Dance
Otaku
Chinese Dance
Jazz
K-Pop

Same Music, Different Reference

Long Video Generation

Case 1
Case 2

Comparison with Baseline Methods

Ablation Study

Dataset

BibTeX

@article{YourPaperKey2024,
  title={Your Paper Title Here},
  author={First Author and Second Author and Third Author},
  journal={Conference/Journal Name},
  year={2024},
  url={https://your-domain.com/your-project-page}
}