2/5/2025 | 11:00 AM - 12:20 PM | Regency B

Super-resolution with latent domain generative models

Abstract

The landscape of image restoration and enhancement is undergoing a dramatic transformation, fueled by the emergence of powerful generative models such as Stable Diffusion and Imagen. These models, originally designed for text-to-image synthesis, are now being adapted for image-to-image tasks, including super-resolution. However, harnessing their full potential for real-world applications presents a unique set of challenges. This talk explores the exciting possibilities and inherent difficulties of productionizing generative models for image enhancement, drawing from our experience at Google in developing cutting-edge super-resolution techniques. We will delve into the intricacies of latent diffusion models, examining key obstacles such as hallucination, artifacts, and computational cost, while highlighting promising pathways for overcoming these hurdles.