 |
Center for Nonlinear Studies |
Diffusion model is a new powerful class of generative model that can synthesize data from noise by modeling the gradient of the log data density (i.e. score function). The resulting generative process resembles sequential denoising from pure noise, and is governed by a simple form of stochastic differential equation (SDE). One can also use diffusion model beyond simple data generation. Specifically, when we have partial observation to the image that we wish to visualize, we can directly utilize the diffusion model as a prior, and iterate data consistency steps to sample from the posterior distribution. I will focus on how inverse problem solving with diffusion models can be performed, and the advantages of such method, giving diverse results ranging from CS-MRI, CT, and computer vision. I will also cover advanced strategies to accelerate diffusion models, and future prospectives.
Host: Hyun Lim (CCS-2) and Marc Klasky (T-5)