Toward Generalizable Deblurring: Leveraging Massive Blur Priors with Linear Attention for Real-World Scenarios

Yuanting Gao1,*,§, Shuo Cao2,3,*, Xiaohui Li2,4,
Yuandong Pu2,4, Yihao Liu2,†,△, Kai Zhang1,†
1Tsinghua University, 2Shanghai AI Laboratory, 3University of Science and Technology of China, 4Shanghai Jiao Tong University
*Equal contribution †Corresponding authors △Project lead
§This work was done during his internship at Shanghai AI Laboratory.
Paper Code arXiv
Overview of GLOWDeblur Framework

(a) Visual comparison on challenging real-world images: our GLOWDeblur effectively restores a wide range of blur patterns, while prior methods often fail in complex scenarios. (b) Quantitative comparison on diverse benchmarks: the left plot shows dataset scores computed by ranking methods on each metric and averaging across metrics; the right plot reports average model scores across all datasets, highlighting the strong generalization ability of GLOWDeblur.

Abstract

Image deblurring has advanced rapidly with deep learning, yet most methods exhibit poor generalization beyond their training datasets, with performance dropping significantly in real-world scenarios. Our analysis shows this limitation stems from two factors: datasets face an inherent trade-off between realism and coverage of diverse blur patterns, and algorithmic designs remain restrictive, as pixel-wise losses drive models toward local detail recovery while overlooking structural and semantic consistency, whereas diffusion-based approaches, though perceptually strong, still fail to generalize when trained on narrow datasets with simplistic strategies. Through systematic investigation, we identify blur pattern diversity as the decisive factor for robust generalization and propose Blur Pattern Pretraining (BPP), which acquires blur priors from simulation datasets and transfers them through joint fine-tuning on real data. We further introduce Motion and Semantic Guidance (MoSeG) to strengthen blur priors under severe degradation, and integrate it into GLOWDeblur, a Generalizable reaL-wOrld lightWeight Deblur model that combines convolution-based pre-reconstruction & domain alignment module with a lightweight diffusion backbone. Extensive experiments on six widely-used benchmarks and two real-world datasets validate our approach, confirming the importance of blur priors for robust generalization and demonstrating that the lightweight design of GLOWDeblur ensures practicality in real-world applications.

The Critical Role of Blur Patterns in Real-World Generalization

Despite recent progress, state-of-the-art deblurring methods struggle to generalize to real-world scenarios, often failing even in visually simple scenes due to a reliance on dataset-specific distributions. Our cross-dataset analysis reveals that blur pattern diversity—rather than just image realism—is the dominant factor in this generalization gap. To address this, we propose Blur Pattern Pretraining (BPP), a strategy that learns intrinsic blur priors from diverse synthetic data. As demonstrated below, BPP effectively bridges distribution gaps and resolves domain conflicts, serving as a core component of our GLOWDeblur framework.

Overview of GLOWDeblur

Overview of GLOWDeblur Method

Overview of GLOWDeblur. The framework integrates a Pre-Reconstruction & Domain-Alignment module with a lightweight diffusion framework, guided by motion maps and cross-modal text semantics. Training involves pre-training on datasets with diverse blur patterns, followed by joint fine-tuning on real-captured datasets.

Quantitative Comparison

Qualitative Comparison

BibTeX

@article{gao2024toward,
  title={Toward Generalizable Deblurring: Leveraging Massive Blur Priors with Linear Attention for Real-World Scenarios},
  author={Gao, Yuanting and Cao, Shuo and Li, Xiaohui and Pu, Yuandong and Liu, Yihao and Zhang, Kai},
  journal={Conference Name},
  year={2024}
}