Abstract
Diffusion models have demonstrated their utility as learned priors forsolving various inverse problems. However, most existing approaches are limitedto linear inverse problems. This paper exploits the efficient and unsupervisedposterior sampling framework of Denoising Diffusion Restoration Models (DDRM)for the solution of nonlinear phase retrieval problem, which requiresreconstructing an image from its noisy intensity-only measurements such asFourier intensity. The approach combines the model-based alternating-projectionmethods with the DDRM to utilize pretrained unconditional diffusion priors forphase retrieval. The performance is demonstrated through both simulations andexperimental data. Results demonstrate the potential of this approach forimproving the alternating-projection methods as well as its limitations.