AC-Motion: High-Quality Arbitrary Motion Style Transfer with Local Channel Modulation and Adaptive Attention
Xinjie Chen, Yuan Ma, Meili Wang · 2025
Motion style transfer plays a crucial role in fields such as animation production and virtual reality. Although recent methods have shown promising results, they often struggle to maintain the structural integrity of content motion when adapting styles that involve motion bobbing behaviors such as jumping, resulting in distortions and unnatural stylized motions. To address these challenges, we propose AC-Motion, a framework comprising two core components: 1) an $\underline{A}$ daptive Attention Network (AdaAtt-Net) that improves local feature representation through a lightweight channel-aware modulation prior to normalization and refines style-content integration via an adaptive attention mechanism; and 2) a Content Consistency Loss, a motionspecific constraint designed to align the stylized motion with the structural characteristics of the source, mitigating unintended body oscillations and subtle structural deviations during style transfer. By jointly leveraging adaptive feature modulation and structure-preserving constraints, AC-Motion generates more natural, accurate stylized motions while maintaining a streamlined architecture. Extensive experiments demonstrate the substantial improvements achieved by our approach across diverse motion style transfer tasks.