In this talk, I will discuss the image priors learned by modern generative models; how to efficiently sample them, how to synthesize gigapixel images, and how to simplify computational photography reconstruction tasks. Early work in our group examined how to utilize the prior knowledge encoded in diffusion models to synthesize samples with unseen conditions, without retraining. Our recent work asks the question of “what makes two diffusion models different apart from image quality” and finds the answer in the properties of the Jacobian of the learned models/priors. Insightful use of these image priors can either lead to generation that is significantly cheaper or generation of images at unprecedented size. We proposed a visual saliency prior to adaptively reduce image and video generation costs. We were the first to synthesize gigapixel-size images by spatially modulating a local prior across a much larger image. Finally, I will introduce a measurement-in-the-loop approach to a computational photography task. In our recent work “PoPPy”, we show how a single polarization measurement can drive, at test time, a learned generative prior of RGB-to-normals for accurate surface normal estimation.

Biography: Professor Dimitris Samaras received the diploma degree in computer science and engineering from the University of Patras, in 1992, the MSc degree from Northeastern University, in 1994, and the PhD degree from the University of Pennsylvania, in 2001. He is a SUNY Empire Innovation Professor of Computer Science with Stony Brook University, where he directs the Computer Vision Lab. His research interests include human behavior analysis, generative models, illumination modeling and estimation for recognition and graphics, and biomedical image analysis. He has co-authored over 250 peer-reviewed research articles in Computer Vision, Machine Learning, Computer Graphics and Medical Imaging conferences and journals. He was Program Chair of CVPR 2022, General Chair of ACCV 2024 and ICCV 2025, and is frequent Area Chair in Computer Vision and Machine Learning conferences.