[论文解读] Toward Generalizable Deblurring: Leveraging Massive Blur Priors with Linear Attention for Real-World Scenarios

Yuan Gao, Shuo Cao+4 more|arXiv (Cornell University)|Jan 10, 2026

Advanced Image Processing Techniques被引用 0

一句话总结

论文提出 GLOWDeblur，通过 Blur Pattern Pretraining (BPP) 学习的模糊先验，并通过 Motion and Semantic Guidance (MoSeG) 进行引导，从而在多样化模糊模式下实现强烈的真实场景泛化。

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/.

研究动机与目标

• Identify why deblurring models fail to generalize to real-world blur patterns beyond training data.
• Quantify the role of blur pattern diversity versus realism in cross-dataset generalization.
• Propose a data-centric pretraining strategy (BPP) to learn blur priors and transfer them to real-world data.
• Develop a lightweight diffusion-based deblurring model (GLOWDeblur) that remains practical for real-world use.
• Enhance priors with motion guidance and cross-modal semantic cues to handle severe blur.

提出的方法

• Blur Pattern Pretraining (BPP) to learn blur priors from large-scale simulated datasets with diverse blur patterns.
• Two-stage training: BPP on simulated data followed by joint fine-tuning on real-world datasets.
• Motion Guidance (MoG) to provide trajectory-based blur cues via a lightweight motion estimator.
• Semantic Guidance (SeG) using cross-modal captions to supply high-level scene semantics to the diffusion backbone.
• GLOWDeblur architecture: a pre-reconstruction & domain-alignment module plus a lightweight diffusion backbone with Deep Compression AutoEncoder and Linear Attention.
• Efficient design choices: SimpleGate activations and Simplified Channel Attention to improve efficiency; latent diffusion in a highly compressed latent space.

实验结果

研究问题

• RQ1 What are the main factors limiting generalization of deblurring models to real-world blur patterns?
• RQ2 How can blur pattern diversity be captured and transferred to improve cross-dataset robustness?
• RQ3 Can a lightweight diffusion-based framework achieve real-world performance while remaining practical for deployment?
• RQ4 Does incorporating motion and semantic guidance further improve restoration under severe blur?

主要发现

• Blur pattern diversity, not just realism, drives cross-dataset generalization gaps in deblurring.
• BPP consistently improves in-distribution accuracy and cross-domain robustness when transferring from simulated to real-world data.
• Naïve mixed-training across datasets degrades performance, whereas BPP bridges distribution gaps.
• GLOWDeblur achieves strong performance across six benchmarks and two real-world datasets, demonstrating improved generalization.
• MoSeG (motion and semantic guidance) reinforces blur priors and aids restoration in severely degraded regions.
• A lightweight diffusion backbone with linear attention and a 32x deep compression autoencoder delivers practical real-world efficiency without sacrificing performance.