Researchers introduce a structure-aware framework for music-driven dance generation that models choreography as a sequence of atomic movements rather than continuous signals.
- The method constructs an atomic movement vocabulary by segmenting large-scale dance data and clustering it into groups.
- A large language model is used to semantically relabel and refine these clusters, creating interpretable and reusable motion events.
- A two-stage generation process mirrors human choreography: first predicting the type, duration, and timing of movements to form a symbolic allocation, then synthesizing smooth motion via a transition-aware generator.
- Experiments show improved structural coherence, rhythmic alignment, and perceptual naturalness compared to baselines, along with enhanced interpretability and controllable editing.
This approach addresses the lack of compositional structure in previous neural methods, making generated dances easier to control and semantically consistent with the input music.