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

GLOWDeblur is a generalizable and lightweight image deblurring framework designed for real-world blur, where many deep learning methods fail to generalize beyond their training distributions.

---

Motivation

Although deep deblurring has advanced rapidly, most methods exhibit poor cross-dataset generalization, with significant performance drops on real captured blur.

Our analysis attributes this to:

• Dataset trade-off: realism vs. coverage of diverse blur patterns
• Algorithmic bias: pixel-wise losses emphasize local details but ignore structural/semantic consistency
• Diffusion limitation: strong perceptual quality yet still brittle when trained on narrow data distributions

Key finding: Blur pattern diversity is the decisive factor for robust real-world generalization.

---

Method

Blur Pattern Pretraining (BPP)

We propose BPP, which learns transferable blur priors from large-scale simulated blur data, then transfers them via joint fine-tuning on real-captured datasets.

Motion & Semantic Guidance (MoSeG)

To improve robustness under severe degradation, MoSeG strengthens blur priors using:

• Motion Guidance (motion estimation / motion maps)
• Semantic Guidance (cross-modal text semantics)

GLOWDeblur

GLOWDeblur combines:

• a convolution-based pre-reconstruction & domain-alignment module
• a lightweight diffusion backbone with linear attention for efficiency

---

Qualitative Results

▶ Click to expand qualitative comparisons

Qualitative comparison on real-world and challenging blur scenarios:

---

BibTeX

@article{gao2026toward,
  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={arXiv preprint arXiv:2601.06525},
  year={2026}
}