<ul data-eligibleForWebStory="true"><li>A new method called ReGuidance has been introduced to improve sample quality and rewards in solving hard inverse problems with low signal-to-noise ratios.</li><li>Diffusion models are being utilized as data priors in solving inverse problems with pretrained models.</li><li>Training-free methods like diffusion posterior sampling (DPS) offer flexible algorithms but struggle on hard inverse problems.</li><li>ReGuidance involves inverting a candidate solution using probability flow ODE and initializing DPS with the resulting latent.</li><li>The ReGuidance wrapper has been tested on tasks like large box in-painting and super-resolution with positive results.</li><li>State-of-the-art baseline methods fail on these tasks while ReGuidance significantly improves sample quality and measurement consistency.</li><li>Theory suggests ReGuidance can enhance rewards and bring solutions closer to the data manifold, especially on multimodal data distributions.</li><li>ReGuidance is the first method to offer a rigorous algorithmic guarantee for diffusion posterior sampling (DPS).</li></ul>

ReGuidance: A Simple Diffusion Wrapper for Boosting Sample Quality on Hard Inverse Problems

Discover more