Calibrated Test-Time Guidance for Bayesian Inference

Researchers have identified flaws in existing test-time guidance methods for diffusion models that prevent proper Bayesian posterior sampling. They propose new estimators that enable calibrated inference, significantly outperforming previous methods on Bayesian tasks and matching state-of-the-art results in black hole image reconstruction.

• →Current test-time guidance methods for diffusion models focus on reward maximization rather than proper Bayesian posterior sampling.
• →Existing approaches produce miscalibrated inference due to structural approximations in their design.
• →New consistent alternative estimators have been developed to enable calibrated sampling from Bayesian posteriors.
• →The proposed methods significantly outperform previous approaches on Bayesian inference tasks.
• →The technique achieves state-of-the-art performance in black hole image reconstruction applications.