Dance Generality
Preserves shared choreographic structure across diverse dance styles.
Abstract
Music-driven 3D dance generation has broad applications in choreography, virtual reality, and creative content creation, yet traditional methods often treat genre as an auxiliary modifier rather than a core semantic driver. This weak genre conditioning can compromise music-motion synchronization and disrupt genre continuity during complex rhythmic transitions.
MEGADance decouples choreographic consistency into dance generality and genre specificity. It combines a High-Fidelity Dance Quantization stage based on Finite Scalar Quantization with a Genre-Aware Dance Generation stage that uses Mixture-of-Experts and a Mamba-Transformer hybrid backbone, achieving strong dance quality and genre controllability on FineDance and AIST++.
Overview Video
Video
Method
Genre-Aware MoE Dance Generation
Genre Specificity
Uses genre-disentangled expert routing to strengthen style continuity.
Music-Motion Alignment
Maps music to dance tokens with a Mamba-Transformer hybrid backbone.
Experiments
Results
Comparison
Four-Way Comparison I
Four-Way Comparison II
Ours vs Lodge
Ours vs FineNet
Ours vs Bailando++
Genre Controllability
Multi-Genre Generation
Ablation
Ablation Overview
Ours vs w/o SE
Ours vs w/o UE
Ours vs w/o Mamba
Citation
BibTeX
@article{tang2026megadance,
title={Megadance: Mixture-of-experts architecture for genre-aware 3d dance generation},
author={Tang, Xulong and Peng, Ziqiao and Hu, Yuxuan and He, Jun and Liu, Hongyan and others},
journal={Advances in Neural Information Processing Systems},
volume={38},
pages={10947--10969},
year={2026}
}