SDPose: Exploiting Diffusion Priors for Out-of-Domain and Robust Pose Estimation
SDPose adapts a pre-trained Stable Diffusion U-Net to estimate human poses from images. By systematically analyzing which parts of the model's internal features are most robust across different visual styles and corruptions, the authors select optimal feature blocks. This leads to a framework that achieves state-of-the-art performance on pose estimation with less training data and better generalization than existing models, making it useful for zero-shot pose guidance in image generation.