Research map / Image generation / Image restoration
ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation
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 ID-Blau, a pixel-wise blur-condition-conditioned diffusion reblurring model that synthesizes diverse blurred images from sharp GoPro images to augment deblurring training, improving four deblurring models.
Problem
Image deblurring has been advanced mainly through architecture design; data augmentation for deblurring is understudied, and GAN-based blur augmentation is uncontrollable and unstable.
Approach
Parameterize blur as a pixel-wise continuous blur condition map (unit motion orientation plus normalized magnitude derived from averaged optical flow) and train a conditional DDPM/DDIM reblurring model on it, so arbitrary blur conditions can be sampled to create unseen blurred training pairs.
Key results
- ID-Blau improves PSNR on GoPro by 0.45 dB on average across four deblurring models, with FFTformer reaching 34.36 dB (+0.15 dB) on GoPro.
- Enhanced Stripformer gains +1.22 dB on RealBlur-R and +1.29 dB on RealBlur-J, the largest gains reported.
- Diffusion-based ID-Blau gives 32.02 dB on MIMO-UNet versus 31.79 dB without diffusion and 31.22 dB baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ID-Blau augmentation to FFTformer | FFTformer · GoPro test | PSNR: +0.15dB | Helped |
| Add ID-Blau augmentation to FFTformer | FFTformer · HIDE | PSNR: +0.32dB | Helped |
| Add ID-Blau augmentation to Stripformer | Stripformer · RealBlur-J | PSNR: +1.29dB | Helped |
| Add ID-Blau augmentation to Stripformer | Stripformer · RealBlur-R | PSNR: +1.22dB | Helped |
| Rotate blur orientations by 30°, 60°, 90°, 120°, or a random angle (ablation) | MIMO-UNet with ID-Blau augmentation · GoPro | PSNR: 31.22 -> 31.99; 31.22 -> 32.00; 31.22 -> 31.97; 31.22 -> 31.95; 31.22 -> 31.96 | Helped |
| Shift the blur-magnitude distribution peak to larger values (ablation) | MIMO-UNet with ID-Blau augmentation · GoPro | PSNR | Hurt |
| Increase the number of ID-Blau augmented samples (ablation) | MIMO-UNet with ID-Blau augmentation · GoPro | PSNR: 0.53dB to 0.87dB | Helped |
| Use ID-Blau augmentation at matched training iterations (ablation) | MIMO-UNet · GoPro | PSNR | Helped |
| Remove the diffusion process from ID-Blau (ablation) | MIMO-UNet with ID-Blau · GoPro | PSNR: 31.22 -> 31.79 | Helped |
| Add the diffusion process to ID-Blau (ablation) | MIMO-UNet with ID-Blau† · GoPro | PSNR: +0.23dB | Helped |
| Add ID-Blau augmentation instead of BGAN augmentation (ablation) | DBGAN · GoPro | PSNR: 30.92 -> 31.57 | Helped |
| Add the diffusion process to ID-Blau (ablation) | DBGAN with ID-Blau† · GoPro | PSNR: 31.57 -> 31.67 | Helped |
Surprising findings
- Shifting the blur magnitude distribution peak to larger values (0.15 to 0.4) decreases deblurring performance, so augmentation must stay near the original distribution.
- Randomly varying blur orientation (specific angles or random) gives similar gains, suggesting specific angles are unnecessary.
Limitations
- Relies on optical-flow estimation (pre-trained network) to compute blur condition maps from sharp sequences, and GoPro's accumulation-based blur as training signal.
- Gains are modest (around 0.5 dB) and evaluated on a limited set of deblurring architectures and datasets.
Experimental setup
Datasets: GoPro HIDE RealBlur-J RealBlur-R
Benchmarks: GoPro HIDE RealBlur-J RealBlur-R
Metrics: PSNR SSIM
Baselines: MIMO-UNet+ Restormer Stripformer FFTformer BGAN DBGAN
Builds on: DDPM DDIM GoPro blur accumulation dataset BGAN
Closest papers by meaning
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.90)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.88)
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.88)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Adaptive Window Pruning for Efficient Local Motion Deblurring (ICLR 2024 · similarity 0.87)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.86)
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.