Solving Bayesian inverse problems with diffusion priors and off-policy RL
This work addresses the problem of effective posterior inference in latent space for researchers in inverse problems, showing that existing training-free methods are biased and incremental improvements are made.
The paper tackled the challenge of performing Bayesian inverse problems using diffusion priors by applying the Relative Trajectory Balance (RTB) off-policy RL objective to train conditional diffusion model posteriors from pretrained unconditional priors, achieving improved performance in linear and non-linear inverse problems in vision and science.
This paper presents a practical application of Relative Trajectory Balance (RTB), a recently introduced off-policy reinforcement learning (RL) objective that can asymptotically solve Bayesian inverse problems optimally. We extend the original work by using RTB to train conditional diffusion model posteriors from pretrained unconditional priors for challenging linear and non-linear inverse problems in vision, and science. We use the objective alongside techniques such as off-policy backtracking exploration to improve training. Importantly, our results show that existing training-free diffusion posterior methods struggle to perform effective posterior inference in latent space due to inherent biases.