Exposing failure modes of diffusion posterior samplers, and mitigating some of them
Diffusion models have shown impressive performance as data-driven Bayesian priors in computational imaging inverse problems, yet are still prone to hallucinating unrealistic images and suppressing uncertainty inherent in ambiguous measurements. In the first part of the talk, we introduce a finite-sample lens onto diffusion posterior sampling that allows us to compare against a known ground-truth posterior distribution and study when, why, and how certain types of posterior samplers err. In the second part of the talk, we propose a simple modification that mitigates one of these failure modes, specifically reducing the incidence of prior-inconsistent (hallucinated) samples. This is joint work with Ben Burns, Merve Kocabas, Andreas Geiger, and Kent Gauen

