Latin dance action style transfer based on improved Ada IN algorithm

This study addresses limitations in current dance action style transfer methods, such as weak spatiotemporal coupling and poor generalisation.It proposes a novel approach using improved adaptive instance normalisation (I Ada IN) with a joint-limb-global layered normalisation structure to enhance style decoupling.The method incorporates a spatiotemporal transformer and inverse kinematics correction to improve stability and style fidelity in long sequences.Experiments show significant gains: a 43% higher style detail retention rate (0.89 vs. 0.62), a 27% improvement in structural similarity (0.94), and a 50% reduction in joint motion error (4.3 mm) over the original Ada IN.With a frame rate of 120 and processing time of 8ms per frame, the model meets real-time performance standards.This method achieves high-fidelity style transfer, accurate content preservation, and stable cross-domain generalisation through innovative hierarchical feature fusion and spatiotemporal modelling strategies, providing feasible technical support and application prospects for virtual dance teaching, intelligent choreography systems, and the digital protection of intangible cultural heritage.