Diffusion-based Event Generation for High-Quality Image Deblurring

Xinan Xie, Qing Zhang, Wei-Shi ZhengCVPR 2025 · OpenReview

Abstract

While event-based deblurring have demonstrated impressive results, they are impractical for consumer photos captured by cell phones and digital cameras that are not equipped with the event sensor. To address this problem, we in this paper propose a novel deblurring framework called Event Generation Deblurring (EGDeblurring), which allows to effectively deblur an image by generating event guidance describing the motion information using a diffusion model. Specifically, we design a motion prior generation diffusion model and a feature extractor to produce prior information beneficial for deblurring, rather than generating the raw event representation. In order to achieve effective fusion of motion prior information with blurry images and produce high-quality results, we develop a regression deblurring network embedded with a dual-attention channel fusion block. Experiments on multiple datasets demonstrate that our method outperforms state-of-the-art image deblurring methods. Our code is available at https://github.com/XinanXie/EGDeblurring.

Introduction

Motion blur is a persistent challenge in photography, occurring when objects move or the camera shakes during exposure. While standard image-only deblurring methods have improved significantly with deep learning, they often struggle with complex, non-linear motion because a single blurry image lacks the temporal history of how the blur occurred. Event cameras, which record pixel-level intensity changes at microsecond resolution, provide this missing temporal information. However, event sensors are not yet found in consumer smartphones or DSLRs. EGDeblurring addresses this gap by using a diffusion model to "hallucinate" dense motion priors from a single blurry image, bringing the power of event-based deblurring to standard cameras.

Methodology

The EGDeblurring framework consists of two primary components and a specialized training strategy:

  1. MPG-Diff (Motion Prior Generation Diffusion Model): Instead of generating raw, sparse "event spikes," this module generates a dense motion prior feature. It is a conditional diffusion model that takes the blurry image and a "detail guidance" map (the high-frequency residuals) as inputs to predict motion information.
  2. RDNet (Regression Deblurring Network): This restoration engine takes the blurry image and the motion prior generated by MPG-Diff to produce the final sharp image.
  3. DACFB (Dual-Attention Channel Fusion Block): A novel module embedded in RDNet that uses cross-modal attention to align and fuse generated motion features with image features, ensuring the hallucinated motion correctly guides pixel restoration.

The system is trained in three stages: first, training RDNet with ground-truth events; second, training MPG-Diff to mimic those features; and finally, jointly fine-tuning both networks to ensure the generated priors are perfectly calibrated for the deblurring task.

Experimental Results

The researchers evaluated EGDeblurring on two major benchmarks: the GoPro dataset (synthetic) and the REBlur dataset (real-world). On the GoPro dataset, EGDeblurring achieved a 34.62 dB PSNR, significantly outperforming other image-only diffusion methods like HI-Diff (33.33 dB) and approaching the performance of methods using actual ground-truth events.

EGDeblurring demonstrates superior restoration of text and license plates, effectively removing ghosting artifacts.

Visual results show that EGDeblurring is particularly good at restoring sharp edges on text and license plates, where traditional methods often leave ghosting artifacts or residual blur. On the real-world REBlur dataset, the method achieved 36.21 dB PSNR, setting a new state-of-the-art for image-only methods.

The framework effectively restores facial details by leveraging hallucinated motion priors.

Conclusion

EGDeblurring effectively creates a "virtual event camera" in software, allowing standard hardware to benefit from temporal-aware architectures originally designed for specialized sensors. By generating feature representations of motion rather than raw events, the framework leverages the strengths of diffusion models while retaining the temporal benefits of event data. Future work may extend this paradigm to other tasks like video frame interpolation or high-dynamic-range (HDR) imaging.

Summary updated

Abstract

While event-based deblurring have demonstrated impressive results, they are impractical for consumer photos captured by cell phones and digital cameras that are not equipped with the event sensor. To address this problem, we in this paper propose a novel deblurring framework called Event Generation Deblurring (EGDeblurring), which allows to effectively deblur an image by generating event guidance describing the motion information using a diffusion model. Specifically, we design a motion prior generation diffusion model and a feature extractor to produce prior information beneficial for deblurring, rather than generating the raw event representation. In order to achieve effective fusion of motion prior information with blurry images and produce high-quality results, we develop a regression deblurring network embedded with a dual-attention channel fusion block. Experiments on multiple datasets demonstrate that our method outperforms state-of-the-art image deblurring methods. Our code is available at https://github.com/XinanXie/EGDeblurring https://github.com/XinanXie/EGDeblurring.
BibTeX
@inproceedings{xie2025diffusion,
  author = {Xinan Xie and Qing Zhang and Wei-Shi Zheng},
  title = {{Diffusion-based Event Generation for High-Quality Image Deblurring}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  url = {https://openreview.net/forum?id=anwJv5SI1k},
}
Figure