AddSR: Accelerating Diffusion-based Blind Super-Resolution with Adversarial Diffusion Distillation

Rui Xie, Ying Tai, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Xiaoqian Ye, qian Wang, Jian Yang

Keywords: Image super-resolution

Abstract:

Blind super-resolution methods based on Stable Diffusion (SD) demonstrate impressive generative capabilities in reconstructing clear, high-resolution (HR) images with intricate details from low-resolution (LR) inputs. However, their practical applicability is often limited by poor efficiency, as they require hundreds to thousands of sampling steps.
Inspired by Adversarial Diffusion Distillation (ADD), we incorporate this approach to design a highly effective and efficient blind super-resolution method. Nonetheless, two challenges arise: First, the original ADD significantly reduces result fidelity, leading to a perception-distortion imbalance. Second, SD-based methods are sensitive to the quality of the conditioning input, while LR images often have complex degradation, which further hinders effectiveness.
To address these issues, we introduce a Timestep-Adaptive ADD (TA-ADD) to mitigate the perception-distortion imbalance caused by the original ADD. Furthermore, we propose a prediction-based self-refinement strategy to estimate HR, which allows for the provision of more high-frequency information without the need for additional modules.
Extensive experiments show that our method,~\name, generates superior restoration results while being significantly faster than previous SD-based state-of-the-art models (e.g., $7\times$ faster than SeeSR).

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6935