Research map / Image generation / Image restoration
BlurDM: A Blur Diffusion Model for 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 BlurDM, a dual noise-and-blur diffusion model that mimics motion-blur formation and is used in latent space as a prior generator to boost existing image deblurring networks.
Problem
Standard diffusion models ignore the physics of motion blur formation (continuous exposure), limiting their usefulness for dynamic scene image deblurring; regression-based deblurring also yields over-smoothed outputs lacking high-frequency detail.
Approach
A forward process that adds both noise and blur residuals (from continuous exposure integration) to a sharp image, and a reverse process that jointly denoises and deblurs starting from the blurred image plus Gaussian noise; implemented in latent space as a prior generator fused into deblurring networks via a Prior Fusion Module with three-stage training.
Key results
- BlurDM improves MIMO-UNet, Stripformer, FFTformer, and LoFormer on GoPro with average PSNR gains of +0.31 dB, and +0.78 dB on RealBlur-J.
- On GoPro, BlurDM with FFTformer reaches 34.34 dB PSNR versus 34.21 dB baseline; with MIMO-UNet it reaches 32.93 vs 32.44 dB.
- BlurDM adds on average 4.16 GFLOPs, 3.33M parameters and about 9 ms inference overhead across the four deblurring models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BlurDM prior generation | MIMO-UNet · GoPro, HIDE, RealBlur-J, RealBlur-R | PSNR: 32.44 -> 32.93; 30.00 -> 30.73; 31.59 -> 32.13; 39.03 -> 39.63 | Helped |
| Add BlurDM prior generation | Stripformer · GoPro, HIDE, RealBlur-J, RealBlur-R | PSNR: 33.09 -> 33.53; 31.03 -> 31.36; 32.48 -> 33.53; 39.84 -> 41.00 | Helped |
| Add BlurDM prior generation | FFTformer · GoPro, HIDE, RealBlur-J, RealBlur-R | PSNR: 34.21 -> 34.34; 31.62 -> 31.76; 32.62 -> 32.92; 40.11 -> 40.55 | Helped |
| Add BlurDM prior generation | LoFormer · GoPro, HIDE, RealBlur-J, RealBlur-R | PSNR: 33.54 -> 33.70; 31.18 -> 31.27; 32.23 -> 33.47; 40.36 -> 40.92 | Helped |
| Add a noise estimator (ablation) | MIMO-UNet baseline · GoPro | PSNR: 31.78 -> 31.91 | Helped |
| Add a blur estimator (ablation) | MIMO-UNet baseline · GoPro | PSNR: 31.78 -> 32.20 | Helped |
| Use both noise and blur estimators (ablation) | MIMO-UNet baseline · GoPro | PSNR: 31.78 -> 32.28 | Helped |
| Replace no prior with an MLP prior (ablation) | MIMO-UNet · GoPro | PSNR: 31.78 -> 31.90 | Helped |
| Replace no prior with a DDPM prior (ablation) | MIMO-UNet · GoPro | PSNR: 31.78 -> 31.91 | Helped |
| Use standard diffusion instead of the MLP prior (ablation) | MIMO-UNet · GoPro | PSNR: 31.90 -> 31.91 | No clear effect |
| Use BlurDM instead of the DDPM prior (ablation) | MIMO-UNet · GoPro | PSNR: 31.91 -> 32.28 | Helped |
| Use BlurDM instead of the RDDM prior (ablation) | MIMO-UNet · GoPro | PSNR: 32.03 -> 32.28 | Helped |
Surprising findings
- Standard DDPM-based prior gives gains comparable to a non-diffusion MLP prior, suggesting the diffusion process alone contributes little.
- The Stage 2 and Stage 3 ablation shows that training without the Stage 1 pre-training yields little gain (31.80 dB vs 31.78 baseline).
Limitations
- Designed for motion blur; may not suit defocus blur, which is depth-dependent.
- Training cost is high: three-stage training takes ~70-141 hours on MIMO-UNet, and no error bars or significance tests are reported.
Experimental setup
Datasets: GoPro HIDE RealBlur-J RealBlur-R
Benchmarks: GoPro HIDE RealBlur-J RealBlur-R
Metrics: PSNR SSIM LPIPS FLOPs parameters
Baselines: MIMO-UNet Stripformer FFTformer LoFormer HI-Diff RDDM DDPM
Builds on: DDPM DDIM RDDM HI-Diff
Closest papers by meaning
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.90)
- Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models (ICLR 2025 · similarity 0.90)
- ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation (CVPR 2024 · similarity 0.90)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Residual Denoising Diffusion Models (CVPR 2024 · similarity 0.89)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.88)
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.