Kinematic‐Aware Diffusion Model Correction for Human Motion Generation via Pose Transfer

This paper presents Kinematic‐Aware Pose Transfer (KAPT), a novel post‐processing approach that corrects kinematic violations in diffusion model outputs by integrating a kinematic validator with pose transfer techniques. When kinematic violations are detected, KAPT leverages pose transfer to map the erroneous pose to a kinematically valid corrected pose. Controlled experiments on six DynamiCtrl‐generated motion sequences demonstrate that KAPT reduces motion jitter in 5 out of 6 videos (e.g., jitter variance decreased by 56.6% on man_1 ), improves acceleration stability and limb‐length consistency in 4 out of 6 videos each, and eliminates all detected biomechanical violations. Against the SmoothNet temporal refiner, KAPT yields statistically significant gains in limb‐length consistency (, paired Wilcoxon, ), a controlled recovery benchmark shows it reduces MPJPE by and restores PCK@0.2 to under severe synthetic degradation, and a sequential KAPT SmoothNet pipeline improves both skeletal consistency and temporal smoothness (). In its present form, KAPT instantiates and enforces upper‐limb (shoulder) kinematic constraints together with limb‐length consistency; we therefore report gains in shoulder‐level kinematic plausibility rather than full‐body physical correctness, and use this more precise term in place of general “kinematic validity” throughout.