Music-to-Dance Generation via Atomic Movements

arXiv cs.AI · 18 hours ago

Researchers developed a framework for generating dance from music by breaking choreography into atomic movements—discrete, semantically meaningful motion units—rather than treating motion as continuous signals. The method uses a large language model to label movement clusters and employs a two-stage process where a model first plans movement sequences timed to music, then generates smooth transitions between them. This approach produces structurally coherent dances with improved rhythmic alignment and enables editing through explicit structural control.