Research map / Image generation / Image restoration
Multiscale Structure Guided Diffusion for Image Deblurring
Mengwei Ren, Mauricio Delbracio, Hossein Talebi, Guido Gerig, Peyman Milanfar
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multiscale structure guidance to image-conditioned diffusion (ablation) | icDPM-L · Realblur-J out-of-domain | LPIPS: 0.156 -> 0.128 | Improved |
| increase icDPM diffusion network size from 6M to 27M without guidance (ablation) | icDPM · Realblur-J out-of-domain | LPIPS: 0.150 -> 0.156 | Hurt |
| increase icDPM diffusion network size from 6M to 27M without guidance (ablation) | icDPM · GoPro in-domain | PSNR: 30.555 -> 32.105 | Improved |
| use smaller guidance network (Guide-S) instead of Guide-L (ablation) | icDPM-L with guidance · Realblur-J out-of-domain | LPIPS: 0.123 -> 0.128 | No clear effect |
| inject guidance in latent space instead of input-level concatenation (ablation) | guidance network · Realblur-J | LPIPS: 0.145 -> 0.143 | Improved |
| use multiscale guidance (x2, x4, x8) instead of single scale (ablation) | guidance network · Realblur-J | LPIPS: 0.143 -> 0.141 | Improved |
| use grayscale regression target instead of RGB (ablation) | guidance network · Realblur-J | LPIPS: 0.141 -> 0.137 | Improved |
| inject guidance by addition instead of concatenation or adaptive norm (ablation) | icDPM · GoPro | PSNR: 30.248 / 29.676 -> 31.139 | Improved |
| add regression loss on guidance features (ablation) | Ours · GoPro | PSNR: 30.56 -> 31.19 | Improved |
| keep input concatenation alongside guidance (ablation) | Ours · GoPro in-domain | PSNR: 25.20 -> 30.65 | Improved |
| add adversarial domain adaptation on guidance features | Ours · Realblur-J | visual quality | Hurt |
| use guided icDPM instead of regression and GAN baselines on average of four test sets | deblurring · Average over GoPro, HIDE, Realblur-J, REDS | LPIPS: 0.133 -> 0.104 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- HINet — HiNet: Deep Image Hiding by Invertible Network (ICCV 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 2025 · similarity 0.92)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.90)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Learning Diffusion Texture Priors for Image Restoration (CVPR 2024 · similarity 0.89)
- Residual Diffusion Deblurring Model for Single Image Defocus Deblurring (AAAI 2025 · similarity 0.89)
- 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.
