Optimizing Diffusion Priors with a Single Observation

The paper introduces a technique to optimize diffusion priors from a single observation, addressing the common issue of priors trained on limited or synthetic data inheriting biases. The method merges multiple diffusion priors into a product-of-experts framework and identifies exponent weights that maximize Bayesian evidence, enabling effective tuning without large observation sets.

Validated on black hole imaging and text-conditioned deblurring, the approach produces more flexible and trustworthy posterior distributions. It avoids overfitting by generalizing priors through exponent weighting, allowing tempered and combined models for better inverse problem solutions.