StableAnimator++ is an ID-preserving video diffusion framework proposed to solve the problem of maintaining ID consistency in existing human image animation diffusion models, especially when the body size or position of the reference image and the driving video are significantly different. It generates high-quality videos based on reference images and pose sequences without post-processing, and is the first model with learnable pose alignment. It performs pose alignment by predicting the similarity transformation matrix between the reference image and the driving pose using a learnable layer based on Singular Value Decomposition (SVD), and improves the face embedding using image and face embedding. In addition, it introduces a distribution-aware ID adapter to prevent interference in the temporal layer and preserves the ID through distribution alignment. In the inference step, it introduces HJB-based face optimization to improve the fidelity of the face during the noise removal process. Its effectiveness is qualitatively and quantitatively proven through benchmark experiments.
