Diffusion-transformer hybrid model for robust Taijiquan motion assessment

Traditional Tai Chi (Taijiquan) assessment suffers from subjectivity and low precision. This study develops a generative artificial intelligence (GAI)-based system integrating High-Resolution Network (HRNet), Vision Transformer for Human Pose Estimation (ViTPose), and diffusion models for intelligent posture evaluation. The framework enables pose reconstruction, motion generation, and semantic scoring. Evaluated on the UMONS-TAICHI dataset, the model achieves Mean Per Joint Position Error (MPJPE) = 7.142, smoothness index = 0.911, and Pose Completion Ratio (PCR) = 96.248. It maintains PCR > 93.957 under 10–40% occlusion, outperforming baseline models. The system accurately evaluates complex movements with coordination and rhythm scores > 0.90. This GAI-driven approach provides an objective, robust tool for Taijiquan digitalization and intelligent training.