Real-SRGD: Enhancing Real-World Image Super-Resolution with Classifier-Free Guided Diffusion

Kenji Doi, Shuntaro Okada, Ryota Yoshihashi, Hirokatsu Kataoka; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 3739-3755

Real-world image super-resolution (RISR) aims to reconstruct high-resolution (HR) images from degraded low-resolution (LR) inputs, addressing challenges such as blurring, noise, and compression artifacts. Unlike conventional super-resolution (SR) approaches that typically generate LR images through synthetic downsampling, RISR confronts the complexity of real-world degradation. To effectively handle the intricate challenges of RISR, we adapt classifier-free guidance (CFG), a technique initially developed for multi-class image generation. Our proposed method, Real-SRGD (Real-world image Super-Resolution with classifier-free Guided Diffusion), decomposes RISR challenges into three distinct sub-tasks: Blind image restoration (BIR), conventional SR, and RISR itself. We then train class-conditional SR diffusion models tailored to these sub-tasks and use CFG to enhance the super-resolution performance in real-world settings. Our empirical results demonstrate that Real-SRGD surpasses existing state-of-the-art methods in both quantitative metrics and qualitative evaluations, as demonstrated by user studies. Moreover, our method demonstrates exceptional generalizability across a range of conventional SR benchmark datasets. The code can be found at https://github.com/yahoojapan/srgd.