Boosting Diffusion Guidance via Learning Degradation-Aware Models for Blind Super Resolution

TL;DR

The paper tackles blind super-resolution under unknown degradation by embedding degradation-aware models into diffusion-guided SR (DADiff) and introducing input perturbation and a guidance scalar to boost performance. It investigates both explicit kernel estimation and implicit degradation modeling, integrating an explicit kernel estimator into the diffusion framework and evaluating its limitations with respect to real-world degradations. A comprehensive data-generation strategy is proposed for DIV2K and CelebA-HQ to train degradation and restoration models, including kernel-based LR synthesis with diverse kernels. Empirical results show degradation-aware diffusion guidance achieves superior performance on blind SR benchmarks, with qualitative demonstrations on ImageNet-Val and CelebA-Val. The work enables higher-fidelity blind SR without requiring known degradation kernels, enhancing applicability in real-world imaging tasks.

Abstract

Recently, diffusion-based blind super-resolution (SR) methods have shown great ability to generate high-resolution images with abundant high-frequency detail, but the detail is often achieved at the expense of fidelity. Meanwhile, another line of research focusing on rectifying the reverse process of diffusion models (i.e., diffusion guidance), has demonstrated the power to generate high-fidelity results for non-blind SR. However, these methods rely on known degradation kernels, making them difficult to apply to blind SR. To address these issues, we present DADiff in this paper. DADiff incorporates degradation-aware models into the diffusion guidance framework, eliminating the need to know degradation kernels. Additionally, we propose two novel techniques: input perturbation and guidance scalar, to further improve our performance. Extensive experimental results show that our proposed method has superior performance over state-of-the-art methods on blind SR benchmarks.

Figures (3)

• Figure 1: Qualitative comparison of $\bf{4 \times}$ upsampling on ImageNet-Val.
• Figure 2: Qualitative comparison of $\bf{4 \times}$ upsampling on CelebA-Val.
• Figure 3: Qualitative comparison of $\bf{8 \times}$ upsampling on CelebA-Val.