Reproduce “Weak Diffusion Priors Can Still Achieve Strong Inverse-Problem Performance” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jing Jia, Wei Yuan, Sifan Liu, Liyue Shen, Guanyang Wang
The authors study whether weak or mismatched diffusion priors can support image restoration. They optimize a generator’s starting noise to match its output to observed measurements. On CelebA face inpainting, a three-step bedroom-trained prior reaches PSNR 32.76, versus 31.98 for full-chain in-domain DPS. This result applies when measurements provide enough anchors; large missing regions can expose failures.
Key insights
- In high-dimensional inverse problems with informative measurements, the posterior distribution can concentrate exponentially on the signal components best matching the observation, reducing the prior's influence.
- Weak diffusion priors, even with low generative fidelity or from mismatched domains, preserve similar short-range local spatial structural information as stronger, matched priors.
- The robustness of weak priors arises from informative measurements providing local constraints, which the prior then propagates through its retained local structural properties to complete unobserved regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.