Toward Generalizable Deblurring: Leveraging Massive Blur Priors with Linear Attention for Real-World Scenarios
Yuanting Gao, Shuo Cao, Xiaohui Li, Yuandong Pu, Yihao Liu, Kai Zhang
TL;DR
This paper introduces a novel deblurring method that leverages diverse blur priors and a lightweight diffusion backbone to improve generalization in real-world scenarios, validated across multiple benchmarks.
Contribution
Proposes Blur Pattern Pretraining and Motion and Semantic Guidance to enhance blur prior transfer, leading to a practical, generalizable lightweight deblurring model called GLOWDeblur.
Findings
GLOWDeblur outperforms existing methods on six benchmarks.
Blur priors significantly improve robustness in real-world deblurring.
The lightweight design ensures practical deployment.
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…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
- The paper is well-motivated. The generalization problem in the image deblurring task is crucial and has not been fully explored. - The characteristic analysis of the current deblurring datasets contributes to the deblurring community. - The paper is well-written and easy to follow.
- There are concerns for the BBP in tackling the generalization problem: BBP relies on GSBlur, a larger existing simulation dataset. Thus, the improvement may be due to 1) GSBlur is a larger training dataset and 2) GSBlur covers the blur characteristics of each test set, rather than generalizing to new blur characteristics. - GLOWDeblur significantly degrades the PSNR and SSIM in most cases in Table 3. However, these are major metrics in the image deblurring task. This method may severely affect
1. Writing: The paper is well-organized and clearly written, making it easy to follow. 2. Logical Idea: The integration of motion-related and semantic-related information for deblurring is intuitive and well-motivated, representing a logical extension of existing approaches. The overall framework design is cohesive, and experiments demonstrate promising results.
1. Motion Guidance: The proposed motion modeling appears limited to 2D directional blur, whereas real-world blur often includes depth-axis motion components. Consequently, BPP may fail to capture full 3D motion complexity, reducing its applicability to realistic scenarios. Moreover, since motion trajectories require paired sharp images to be computed, the motion guidance component can only be trained on synthetic datasets, potentially restricting its generalization to real-world data. 2. Semant
1. Performance shows that the proposed training strategy is helpful. 2. The lightweight diffusion architecture is novel, which is a good tradeoff between generalization and efficiency,
1. I think the paper should specifies that it is proposed to handling motion blur instead of general blurs. 2. The training pipeline is complex and hard to reproduce. 3. No explicit inference time evaluation.
1. Precise Problem Definition with Rigorous Evidence: The paper's motivation is clear and highly persuasive. It moves beyond the conventional discussion of "realism" by skillfully combining Table 1 (performance degradation) and Figure 3 (imbalanced pattern distributions). This robustly proves that the root cause of generalization failure is the biased and imbalanced distribution of "blur patterns" in the training data. 2. Shifting Research Focus with a Systematic Strategy: Based on this key
1. Weak Methodological Innovation and Unaddressed Concerns: The paper's novelty is limited, primarily relying on stacking existing modules like MoG and SeG. These additions also introduce significant concerns: the accuracy of MoG's motion estimation is unverified and risks misguiding the restoration. Furthermore, the usage of SeG is ambiguous; if a VLM is required at test time, it introduces an unfair external annotation, and its robustness on low-quality, poorly-described images is unexplored.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Image Processing Techniques · Generative Adversarial Networks and Image Synthesis · Image and Video Quality Assessment