Training-Free Diverse Sampling via Local Pullback Geometry

This method improves diversity in conditional diffusion models by steering sampling trajectories along directions of maximal output sensitivity. By leveraging the local pullback geometry between embedding and image spaces, it encourages distinct generations without additional optimization or training.