Weak Diffusion Priors Can Still Achieve Strong Inverse-Problem Performance
This addresses the practical challenge of applying diffusion models to inverse problems without needing high-fidelity training data, though it is incremental in providing theoretical justification and empirical analysis.
The paper tackles the problem of using mismatched or low-fidelity diffusion priors for inverse problems, finding that they often perform nearly as well as in-domain baselines when measurements are highly informative, such as with many observed pixels, and identifies failure regimes.
Can a diffusion model trained on bedrooms recover human faces? Diffusion models are widely used as priors for inverse problems, but standard approaches usually assume a high-fidelity model trained on data that closely match the unknown signal. In practice, one often must use a mismatched or low-fidelity diffusion prior. Surprisingly, these weak priors often perform nearly as well as full-strength, in-domain baselines. We study when and why inverse solvers are robust to weak diffusion priors. Through extensive experiments, we find that weak priors succeed when measurements are highly informative (e.g., many observed pixels), and we identify regimes where they fail. Our theory, based on Bayesian consistency, gives conditions under which high-dimensional measurements make the posterior concentrate near the true signal. These results provide a principled justification on when weak diffusion priors can be used reliably.