CoheDancers: Enhancing Interactive Group Dance Generation through Music-Driven Coherence Decomposition

TL;DR

CoheDancers addresses the coherence gap in Group Music2Dance by decomposing the problem into synchronization, naturalness, and fluidity. It introduces Cycle Consistency based Dance Synchronization, Auto-Regressive Exposure Bias Correction, and Adversarial Training to enforce robust music-dance correspondences, realistic motion, and natural group dynamics. The authors also present I-Dancers, a diverse, open-source dataset with 3.8–3.9 hours of motion and music across 12 genres, plus evaluation metrics for global semantic and local synchronization. Experiments on I-Dancers and AIOZ-GDANCE demonstrate state-of-the-art performance and strong qualitative results, with plans to release code publicly.

Abstract

Dance generation is crucial and challenging, particularly in domains like dance performance and virtual gaming. In the current body of literature, most methodologies focus on Solo Music2Dance. While there are efforts directed towards Group Music2Dance, these often suffer from a lack of coherence, resulting in aesthetically poor dance performances. Thus, we introduce CoheDancers, a novel framework for Music-Driven Interactive Group Dance Generation. CoheDancers aims to enhance group dance generation coherence by decomposing it into three key aspects: synchronization, naturalness, and fluidity. Correspondingly, we develop a Cycle Consistency based Dance Synchronization strategy to foster music-dance correspondences, an Auto-Regressive-based Exposure Bias Correction strategy to enhance the fluidity of the generated dances, and an Adversarial Training Strategy to augment the naturalness of the group dance output. Collectively, these strategies enable CohdeDancers to produce highly coherent group dances with superior quality. Furthermore, to establish better benchmarks for Group Music2Dance, we construct the most diverse and comprehensive open-source dataset to date, I-Dancers, featuring rich dancer interactions, and create comprehensive evaluation metrics. Experimental evaluations on I-Dancers and other extant datasets substantiate that CoheDancers achieves unprecedented state-of-the-art performance. Code will be released.

Figures (5)

• Figure 1: CoheDancers synergistically utilizes a Cycle Consistency-based Dance Synchronization strategy, adversarial training, and exposure bias-corrected autoregression to enhance coherence.
• Figure 2: Pipeline Structure. CoheDancers comprises Music2Dance Generation block ($G_{M2D}$), Dance2Music Generation block ($G_{D2M}$), and Strategy Training block ($ST^{M}, ST^{D}$), working synergistically to enhance dance coherence.
• Figure 3: Strategy Training. Strategy comprises the (a) Cycle Consistency based Dance Synchronization strategy, (b) the Adversarial Training pipeline, and (c) the Auto-Regressive-based Exposure Bias Correction strategy.
• Figure 4: Dance sequence examples from I-Dancers dataset.
• Figure 5: Dance sequences generated by CoheDancers on I-Dancers, AIOZ-GDANCE-P2 and AIOZ-GDANCE-P3, from top to bottom.