Music-to-Dance Generation via Atomic Movements

TL;DR — A structure-aware music-to-dance generation framework that models choreography as sequences of interpretable "atomic movements" rather than continuous motion, improving coherence, control, and editability. It matters because it brings compositional structure and interpretability to cross-modal (music→motion) generation.

• Builds an atomic-movement vocabulary by segmenting and clustering large-scale dance data, then using an LLM to semantically relabel/refine clusters into reusable, interpretable motion events.
• Uses a two-stage pipeline mirroring human choreography: a planning stage predicts atomic-movement type, duration, and timing conditioned on music (symbolic allocation), then a transition-aware generator synthesizes smooth motion.
• Reported gains over baselines in structural coherence, rhythmic alignment, and perceptual naturalness, plus controllable editing via explicit structural representation.
• Note: takeaways are drawn from the abstract only; no specific quantitative metrics or datasets are provided in the content.