Research map / Image generation / Image restoration
Residual Diffusion Deblurring Model for Single Image Defocus 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 RDDM, a residual conditional diffusion model for single-image defocus deblurring that uses a pre-deblur module guided by a defocus map and a few-step (2-4) DDIM denoiser built from RDD blocks.
Problem
Single image defocus deblurring struggles with large, spatially varying defocus blur; existing regression methods give poor perceptual quality and generalize poorly, while perceptual losses overfit.
Approach
Formulates defocus deblurring as conditional diffusion that models residuals on top of a cheap pre-deblurred estimate, conditioned on a pretrained defocus map and multi-scale semantic guidance, with deformable-convolution RDD blocks giving learnable receptive fields, enabling only 2-4 inference steps.
Key results
- On DPDD, RDDM reports LPIPS 0.166 vs 0.183 for DRBNet (best prior), PSNR 25.97, SSIM 0.811, outperforming all compared methods.
- On RealDOF (generalization, trained only on DPDD), RDDM achieves PSNR 25.03, SSIM 0.772, and best MAE 0.038 and LPIPS 0.271, ahead of IFANet's LPIPS 0.304.
- Uses only 2 inference steps for RDDM in the table versus 10 for DiffLL and 4 for DiffIR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RDDM instead of DRBNet | single-image defocus deblurring · DPDD test | LPIPS: 0.183 -> 0.166 | Helped |
| Use RDDM instead of DRBNet | single-image defocus deblurring · DPDD test | MAE: 0.039 -> 0.037 | Helped |
| Use RDDM instead of DPDNet | single-image defocus deblurring · DPDD test | FID: 2.112 -> 1.681 | Helped |
| Use RDDM instead of DRBNet | single-image defocus deblurring · DPDD test | PSNR: 25.73 -> 25.97 | Helped |
| Use RDDM instead of DRBNet | single-image defocus deblurring · DPDD test | SSIM: 0.791 -> 0.811 | Helped |
| Use RDDM instead of IFANet | single-image defocus deblurring · RealDOF | LPIPS: 0.304 -> 0.271 | Helped |
| Use RDDM instead of DRBNet | single-image defocus deblurring · RealDOF | MAE: 0.036 -> 0.038 | Hurt |
| Use RDDM instead of IFANet | single-image defocus deblurring · RealDOF | FID: 2.706 -> 2.912 | Hurt |
| Use RDDM instead of IFANet | single-image defocus deblurring · RealDOF | PSNR: 24.71 -> 25.03 | Helped |
| Use RDDM instead of IFANet | single-image defocus deblurring · RealDOF | SSIM: 0.763 -> 0.772 | Helped |
Surprising findings
- On RealDOF, RDDM's FID (2.912) trails IFANet (2.706), so it is not best on all perceptual metrics.
Limitations
- Relies on a pretrained defocus map estimator trained on a separate real-world blur-map dataset.
- Evaluation limited to DPDD, RealDOF and CUHK; no quantitative results on CUHK reported.
Experimental setup
Datasets: DPDD CUHK blur detection dataset RealDOF Lee et al. 2019 defocus blur map dataset
Benchmarks: DPDD RealDOF CUHK
Metrics: LPIPS MAE FID PSNR SSIM
Baselines: DPDNet KPAC IFANet DRBNet DiffLL DiffIR
Builds on: NAFNet DDPM DDIM Restoration with residual diffusion (Whang et al. 2022) DPDNet
Closest papers by meaning
- Prior and Prediction Inverse Kernel Transformer for Single Image Defocus Deblurring (AAAI 2024 · similarity 0.91)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.90)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Quad-Pixel Image Defocus Deblurring: A New Benchmark and Model (CVPR 2025 · similarity 0.89)
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.88)
- ID-Blau: Image Deblurring by Implicit Diffusion-based reBLurring AUgmentation (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.