MeanFlow Few-Step Generation
Predicts interval-averaged velocity for stable, high-fidelity sampling.
Abstract
Music-to-dance generation translates auditory signals into expressive human motion, yet existing approaches still struggle to balance refined 3D motion quality with strict inference budgets. FlowerDance is designed for both physically plausible, artistically expressive motion and efficient generation in speed and memory usage.
FlowerDance combines MeanFlow with Physical Consistency Constraints for high-quality few-step sampling, and uses a lightweight non-autoregressive BiMamba backbone with Channel-Level Fusion for long-horizon music-to-dance synthesis. It also supports motion editing through time-decayed soft masking, enabling users to refine generated dance sequences interactively.
Overview Video
Video
Method
FlowerDance Framework
Physical Consistency Constraints
Regularizes reconstructed motion, velocity, and 3D joints toward plausible kinematics.
BiMamba Channel-Level Fusion
Uses a lightweight non-autoregressive backbone for long-horizon music-motion alignment.
Efficiency
Fast Long-Horizon Generation
63M
Params
4145
FPS
Results
Music-to-Dance Results
Comparison
Ours vs MEGA
Ours vs Lodge
Ours vs FineNet
Four-Way Comparison I
Four-Way Comparison II
Ablation
Generative Model
Model Backbone
Channel-Level Fusion
Physical Consistency Constraint
MeanFlow Loss
Sampling Strategy
ODE Sampling Step
Motion Editing
Citation
BibTeX
@inproceedings{yang2026flowerdance,
title={FlowerDance: MeanFlow for Efficient and Refined 3D Dance Generation},
author={Yang, Kaixing and Tang, Xulong and Peng, Ziqiao and Zhang, Xiangyue and Chen, Chubin and Zhou, Xukun and Wang, Puwei and Liu, Hongyan and He, Jun},
booktitle={European Conference on Computer Vision (ECCV)},
year={2026}
}