Reproduce “Ambient Diffusion Posterior Sampling: Solving Inverse Problems with Diffusion Models Trained on Corrupted Data” on OpenResearch

About this paper

Asad Aali, Giannis Daras, Brett Levac, Sidharth Kumar, Alexandros G. Dimakis, Jonathan I. Tamir

Researchers at Stanford University, MIT CSAIL, UT Austin (IFML), UC Berkeley, and Bespoke Labs introduced Ambient Diffusion Posterior Sampling (A-DPS), a framework that allows diffusion models trained exclusively on linearly corrupted data to solve diverse linear inverse problems. The framework demonstrates that these models can achieve superior reconstruction quality and faster inference, particularly in highly ill-posed scenarios like high-acceleration MRI reconstruction and cross-corruption natural image restoration, when compared to models trained on pristine data.

Key insights

• Diffusion models trained solely on linearly corrupted data can learn robust generative priors that are highly effective for solving inverse problems, even when the inference task involves different corruption types or higher corruption levels.
• A "performance cross-over" phenomenon was observed, where models trained on corrupted data consistently outperform clean-data-trained models in high-corruption inference regimes.
• Ambiently trained diffusion models demonstrate computational efficiency, achieving competitive reconstruction quality with a significantly lower number of function evaluations (NFEs) compared to models trained on clean data.

Summary by alphaXiv.