AffectMoE: continuous valence–arousal soft routing of mixture of experts for emotion-consistent 3D dance generation

Music-driven 3D dance generation requires the affective character of music to be reflected in the generated movements, yet prevailing approaches treat emotion as a shallow auxiliary condition, leading to mismatches between the affective character of music and the expressiveness of generated movements. This study presents AffectMoE, a diffusion-based framework that introduces a continuous valence–arousal soft routing mechanism over a mixture of emotion-specific and emotion-agnostic experts. Unlike prior methods that rely on discrete genre labels or single-module emotional alignment, AffectMoE employs learnable emotion prototype experts distributed across the VA space, with a router producing convex combination weights that enable smooth interpolation of motion dynamics along continuous affective dimensions. Cross-modal contrastive alignment and emotion consistency regularization are integrated with weak load balancing to prevent expert collapse. Experiments on FineDance and AIST++ show an FIDk of 47.8 with statistically significant improvements and a quadrant accuracy of 73% under the proposed Emotion Consistency Score. Continuous soft routing produces monotonic controllability curves in override mode, which single-module alignment does not exhibit under the same setting. A 30-participant user study indicates higher perceived emotional congruence for the proposed approach.