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

TL;DR

This work tackles the problem of poor real-world generalization in image deblurring by identifying blur-pattern diversity as a key factor and dataset bias as a major contributor to cross-dataset gaps. It introduces Blur Pattern Pretraining (BPP) to learn robust blur priors from simulated data, and Motion and Semantic Guidance (MoSeG) to reinforce priors with motion cues and high-level semantics, all integrated into a lightweight diffusion-based framework called GLOWDeblur. GLOWDeblur combines a Pre-Reconstruction & Domain-Alignment module, a Deep Compression AutoEncoder, and a Linear Diffusion Transformer with linear attention to balance restoration quality and real-world efficiency, achieving strong generalization across six benchmarks and two real-world datasets. The results demonstrate that leveraging blur priors and guidance signals enables practical, robust deblurring suitable for real-world applications, with significant efficiency gains over traditional diffusion models.

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 project page is available at https://vegdog007.github.io/GLOWDeblur_Website/.

Figures (14)

• Figure 1: (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.
• Figure 2: Challenges for Real-World Generalization
• Figure 3: Illustration of dataset-specific blur patterns, highlighting notable distribution differences.
• Figure 4: 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.
• Figure 5: Qualitative comparison on GoPro, BSD, and RSBlur (From top to bottom). GLOWDeblur effectively handles diverse blur patterns with high-quality restorations.