Atomic Movements for Music-to-Dance

A structure-aware music-to-dance framework plans interpretable atomic movements before synthesizing smooth continuous dance motion.

TL;DR — The abstract proposes a structure-aware music-to-dance framework that represents choreography as sequences of semantically interpretable atomic movements. It builds an atomic movement vocabulary by segmenting and clustering large-scale dance data, then using a large language model to relabel and refine the clusters. Generation has two stages: predict movement type, duration, and timing from music to form a symbolic dance allocation, then synthesize continuous motion with a transition-aware generator. The abstract claims improved structural coherence, rhythmic alignment, perceptual naturalness, interpretability, and controllable editing, but provides no numerical results, named datasets, authors, arXiv ID, URL, or named baselines.

Problem: dance needs compositional structure

Evidence: "Music-driven dance generation aims to produce human motion that is both rhythmically synchronized and semantically consistent with music." The problem is not merely to create realistic human motion, but to generate dance that fits the music in rhythm and meaning.

The abstract says recent neural methods have made progress in visual realism, but criticizes them for treating motion mainly as a continuous signal. Evidence: "they typically model motion as a continuous signal and neglect its compositional nature." According to the abstract, this makes generated dances "structurally incoherent and difficult to control."

The proposed remedy is to add an explicit structural layer: choreography is represented as a sequence of atomic movement events before final motion is synthesized.

Full analysis, extracted claims, numerical results, entity graph, FAQ, related work, applications, and BibTeX are available via x402 micropayment.