Adaptive Destruction Processes for Diffusion Samplers

In this paper, we present the benefits of using learnable variance and learnable destruction process in diffusion samplers. We empirically find that these modifications help to more accurately model complex energy landscapes, especially with few sampling steps. We also contribute to the understanding of training diffusion samplers by studying techniques that improve their stability and convergence speed. We hope that these results inspire the community to scale our findings to other distributions and domains. Another interesting direction for future work would be to rigorously study the optimal parametrizations of the generation and destruction processes -- including non-Gaussian transitions -- and the theoretical limits of sampling with discrete-time learned diffusions.