Atomic Dance: Explainable Music-to-Dance Generation Using Atomic Movements

Problem Motivation

Current music‑to‑dance models treat generation as an end‑to‑end frame‑wise pose prediction. Although they capture local rhythm, they ignore the hierarchical structure of choreography, leading to long‑duration inconsistency, repetitive motions without variation, and limited editability.

Atomic Movement Definition

An Atomic Movement is defined by three criteria:

Completeness: the unit covers the entire motion process (e.g., a kick includes preparation, thrust, and retraction).

Repeatability with variation: the same unit may appear multiple times with changes in speed, direction, amplitude, or force.

Clear semantics: each unit can be described in natural language such as “raise right hand” or “spin torso”.

Atomic Library Construction Pipeline

The library is built through a three‑stage pipeline (Fig. 1):

Segmentation: continuous dance recordings are split into full‑action events rather than fixed‑length clips.

Clustering: motion features are used to group repeatedly occurring basic patterns.

LLM semantic re‑clustering: a large language model refines each cluster by assigning fine‑grained natural‑language labels.

The resulting atomic library is repeatable, composable, and interpretable, providing the foundation for structured generation.

Two‑Stage Generation Framework

Motion Planning: the model ingests the entire music piece, predicts the temporal positions, types, and durations of atomic movements, and produces a high‑level “dance score”.

Dance Completion: for each planned unit, a prototype is selected from the atomic library; a diffusion model synthesizes smooth transitions and introduces variation for repeated units, preserving overall structure while enriching expression.

This design mirrors real‑world choreography: a choreographer plans the structure, and a dancer executes the motions. Compared with end‑to‑end generators, the framework leverages full‑track music information, yields dances that align better with musical structure, and enables direct editing of atomic units.

Quantitative Evaluation

On the AIST++ dataset, AtomicDance achieves the best scores among state‑of‑the‑art methods on:

Realism (FID)

Rhythm alignment (Beat Alignment Score)

Structural consistency (R‑precision)

Bold numbers in the original table indicate the top performance.

Ablation Studies

Removing the atomic‑movement recreation step degrades realism, rhythm alignment, and structural consistency, confirming its contribution (Table 2). Different atomic‑movement selection strategies were compared; a length‑based selection scheme yielded the highest overall performance (Table 3).

Qualitative Results

Visual comparisons (Fig. 3) and accompanying videos demonstrate clearer motion continuity, stronger musical structure adherence, and more natural transitions than prior methods.

Impact of the Atomic Representation

Introducing an intermediate atomic representation enables the model to learn explicit choreography rules, improving interpretability, editability, and controllability of generated dances.

Potential Applications

The approach opens pathways for digital‑human performance, intelligent choreography tools, human‑AI collaborative art creation, and broader research on structured motion generation and explainable generative models.

Resources

Paper: https://arxiv.org/abs/2607.13978 Code repository: https://github.com/oceanflowlab/AtomicDance Project page:

https://cxhcmhhh.github.io/AtomicDanceProject/