Compositional Text Conditioning for Diffusion-Based Human Motion Generation

Text-to-motion diffusion models typically encode natural language descriptions into a single holistic embedding, limiting controllability for motions involving multiple semantic components. We propose a compositional text-conditioning framework that factorizes semantics into upper- and lower-body components and jointly conditions the diffusion process. Compositional motion － language supervision is constructed via joint-level masking and recombination while preserving full-body coherence. Experiments on a compositional extension of the HumanML3D dataset show improved higher-rank R-Precision over text concatenation, while maintaining comparable motion diversity with a moderate increase in Fréchet Inception Distance (FID).