SPH-DiffDance: Semantic Prototype Reconstruction and Hypergraph-Enhanced Diffusion Framework for Long-Sequence Music-Driven Dance Generation
Year: 2026
Type: preprint
Abstract: Long-sequence music-driven dance generation remains a challenging task, as existing methods still struggle with unstable global choreography semantics, insufficient local pose coordination, and difficulty in preserving complex motion details. Aiming to address these issues, we propose a Semantic Prototype Reconstruction and Hypergraph-Enhanced Diffusion Framework for Long-Sequence Music-Driven Dance Generation (SPH-DiffDance), refining motion from global semantic structure to local pose details. SPH-DiffDance follows a coarse-to fine diffusion framework, designing a Semantic Prototype Reconstruction (SPR) module at the coarse stage to provide semantic guidance for motion generation through constructing a learnable global dance prototype library. The motion latent features are queried and reconstructed, transforming continuous motion representations into prototype representations with semantic constraints. This process enhances the thematic consistency and long-range choreography stability of generated dances. Building upon established global semantics, a Hypergraph-Enhanced Pose Refinement (HEPR) module further improves local motion quality by modeling high-order motion correlations among the trunk, limbs, and feet through hyperedges. Compared with conventional skeleton graphs that mainly describe pairwise joint connections, the hypergraph structure provides richer representations of complex human motion relationships and improves pose coordination and structural rationality. Foot-contact constraints are also incorporated to refine local motion details and reduce foot sliding and abnormal poses. Experiments on the AIST++ music-driven dance generation dataset demonstrate that SPH-DiffDance can generate long-sequence 3D dances with good musical synchronization, motion continuity, and structural consistency.
Source: Research Square
Authors: Zhengping Hu, Yifan Wu, Xingyu Li, Chuanxin Xu
Institution: Yanshan University
Language: English
Topic: Human Motion and Animation
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Awards: G2009310441, G2772709459
OpenAlex ID: W7225390325