Research map / Image generation / Image restoration
Diffusion-based Event Generation for High-Quality Image Deblurring
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EGDeblurring, which uses a conditional diffusion model (MPG-Diff) to generate motion-prior features from a single blurry image and a regression network (RDNet) with a dual-attention channel fusion block to deblur it without event input.
Problem
Event-based deblurring achieves strong results but requires event streams at inference, which consumer cameras lack; image-only deblurring struggles with complex motion blur.
Approach
Instead of generating raw sparse events, a diffusion model generates dense event-derived motion prior features conditioned on the blurry image and a detail guidance map; a regression network fuses these via a dual-attention channel fusion block, trained in three stages.
Key results
- On GoPro, the full image-only EGDeblurring reaches 34.62 dB PSNR / 0.973 SSIM, above AdaRevD (34.60 dB) and HI-Diff (33.33 dB).
- On REBlur, the image-only model reaches 36.21 dB PSNR, a 0.3 dB gain over the best image-only baseline cited in the paper's table.
- Stage-1 model with ground-truth events reaches 35.50 dB on GoPro, above the event-based EFNet (35.46 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EGDeblurring instead of AdaRevD | Image-only deblurring on GoPro · GoPro | PSNR: 34.60 -> 34.62 | Helped |
| Use EGDeblurring instead of HI-Diff | Image-only deblurring on REBlur · REBlur | PSNR: 35.91 -> 36.21 | Helped |
| Add the event-trained Stage 1 model instead of EFNet | Event-based deblurring on GoPro · GoPro | PSNR: 35.46 -> 35.50 | Helped |
| Add the event-trained Stage 1 model instead of REFID | Event-based deblurring on REBlur · REBlur | PSNR: 38.34 -> 38.51 | Helped |
| Generate dense motion-prior features rather than raw events (ablation) | EGDeblurring · GoPro | PSNR: 32.33 -> 34.62 | Helped |
| Use MPG-Diff rather than a convolutional neural network for prior generation (ablation) | EGDeblurring · GoPro | PSNR: 32.47 -> 34.62 | Helped |
| Add DACFB instead of feature concatenation and convolutional layers (ablation) | EGDeblurring RDNet · GoPro | PSNR: 33.91 -> 34.62 | Helped |
| Add detail guidance to the diffusion model (ablation) | EGDeblurring · GoPro | PSNR: 33.84 -> 34.62 | Helped |
| Use 2 value tensors instead of 1 in DACFB (ablation) | EGDeblurring DACFB · GoPro | PSNR: 34.10 -> 34.62 | Helped |
| Use 2 value tensors instead of 3 in DACFB (ablation) | EGDeblurring DACFB · GoPro | PSNR: 34.61 -> 34.62 | Helped |
| Use 2 value tensors instead of 4 in DACFB (ablation) | EGDeblurring DACFB · GoPro | PSNR: 34.61 -> 34.62 | Helped |
| Use 2 value tensors instead of 5 in DACFB (ablation) | EGDeblurring DACFB · GoPro | PSNR: 34.59 -> 34.62 | Helped |
Surprising findings
- Generating raw event representations with the diffusion model performs worse than generating dense motion-prior features (w/ raw event: 32.33 dB vs 34.62 dB on GoPro).
Limitations
- Requires ground-truth events during training, so the event feature extractor and paired event data are still needed.
- Gains over strong image-only methods such as AdaRevD are modest on GoPro.
Experimental setup
Datasets: GoPro REBlur
Benchmarks: GoPro REBlur
Metrics: PSNR SSIM
Baselines: HI-Diff AdaRevD Restormer FFTformer Stripformer HINet EFNet EDI REFID eSL-Net
Builds on: DDPM DDIM EFNet HI-Diff
Closest papers by meaning
- Frequency-aware Event-based Video Deblurring for Real-World Motion Blur (CVPR 2024 · similarity 0.88)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.87)
- DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Deblurring 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
- Motion Deblurring via Spatial-Temporal Collaboration of Frames and Events (AAAI 2024 · similarity 0.86)
- Asymmetric Hierarchical Difference-aware Interaction Network for Event-guided Motion Deblurring (AAAI 2025 · similarity 0.85)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.