אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

Denoising Diffusion Probabilistic Models (DDPM), also known as diffusion models, have recently emerged as state-of-the-art generative models, synthesizing images with unprecedented quality and realism. At their core, diffusion models employ an MSE-trained denoiser neural network in an iterative scheme, transforming random noise into pristine images. Theoretically, this algorithm is proven to draw samples from a learned prior image distribution. In our work, we adapt pre-trained diffusion models for the task of image restoration, mainly focusing on linear inverse problems. We present a novel outlook on inverse problem solving, posing it as a posterior sampling task rather than an optimization problem. This approach introduces several advantages such as improved perceptual quality, multiple solutions, and uncertainty quantification. Moreover, our method does not require task-specific training, and we demonstrate its use for inpainting, super resolution, deblurring, colorization, and more.