Exploiting diffusion prior for image super-resolution

Abstract

Image super-resolution refers to the fundamental task in computer vision that produces high-resolution images from low-resolution inputs. This process has diverse applications, including medical imaging, surveillance and security, satellite imagery and scientific research. Traditional approaches for super-resolution are often constrained by predefined assumptions about the image, whereas recently developed deep generative models have overcome these limitations, demonstrating significant improvements in realism and versatility in handling different scenarios.

In recent years, diffusion models have shown a remarkable ability to generate highly detailed and realistic images, often outperforming the GANs-based generative models. This motivates the incorporation of diffusion models into the deep image priors, aiming to further improve the accuracy of the results.

This paper presents a new approach based on Deep Generative Prior (DGP) for super-resolution tasks by incorporating DGP with diffusion models. More precisely, we integrate the open-source text-to-image generation model Stable Diffusion 3 into DGP models to capture richer image statistics and improve super-resolution outputs. We compare the performance of the integrated solution with a GANs-based approach and demonstrate that diffusion models can effectively generate sharp high-frequency details and solve the one-to-many problem for super-resolution.