Research map / Image generation / Image restoration
Taming Generative Diffusion Prior for Universal Blind Image Restoration
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 BIR-D, a diffusion-based universal blind image restoration method that uses an optimizable convolutional kernel and mask as a learned degradation model and an empirical adaptive guidance scale, evaluated on faces, linear inverse problems, low-light, HDR, and motion deblurring.
Problem
Blind image restoration with unknown degradation functions and parameters; existing diffusion-based restoration methods assume the degradation type and use a manually tuned, constant guidance scale.
Approach
Model the unknown degradation as an optimizable convolutional kernel plus a mask, updated by gradient descent at each reverse diffusion step against the degraded input, and set the guidance scale per step with a derived empirical formula instead of grid search.
Key results
- On LFW and WIDER blind face restoration, BIR-D reaches NIQE 3.94 and 3.65, best among compared methods, with FID 31.49 on WIDER.
- On ImageNet 4x super-resolution, BIR-D achieves the best Consistency (6.32) and FID (37.54) among compared methods.
- On LOL low-light enhancement, BIR-D reports PSNR 14.52, LOE 105.42, FID 68.98 and PI 4.87, outperforming GDP and Zero-DCE on LOE, FID and PI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add adaptive guidance scale (ablation) | BIR-D · LOL low-light enhancement | FID: 82.14 -> 68.98 | Helped |
| Add adaptive guidance scale (ablation) | BIR-D · LOL low-light enhancement | PSNR: 14.35 -> 14.52 | Helped |
| Dynamically update the convolution kernel (ablation) | BIR-D · LOL low-light enhancement | FID: 102.47 -> 68.98 | Helped |
| Dynamically update the convolution kernel (ablation) | BIR-D · LoLi-Phone low-light enhancement | LOE: 102.55 -> 72.83 | Helped |
| Use a 5 × 5 kernel instead of kernel size=7 (ablation) | BIR-D · Low-light enhancement | FID: 70.55 -> 68.98 | Helped |
| Use a 5 × 5 kernel instead of kernel size=7 (ablation) | BIR-D · Motion blur reduction | PSNR: 33.94 -> 34.12 | Helped |
| Use a 9 × 9 kernel instead of kernel size=11 (ablation) | BIR-D · ImageNet-1k 4× super-resolution | PSNR: 24.36 -> 24.58 | Helped |
| Use a 9 × 9 kernel instead of kernel size=11 (ablation) | BIR-D · ImageNet-1k deblur | PSNR: 26.79 -> 27.14 | Helped |
| Use the first-stage pre-training model (ablation) | BIR-D · ImageNet-1k deblur, random initial value | PSNR: 25.88 -> 26.31 | Helped |
| Use the first-stage pre-training model (ablation) | BIR-D · ImageNet-1k deblur, biased initial value | PSNR: 21.49 -> 25.97 | Helped |
Surprising findings
- Ablation shows that PSNR and SSIM on LOL are not the best for BIR-D; Zero-DCE and ExCNet achieve higher PSNR/SSIM than BIR-D on LOL and VE-LOL-L.
Limitations
- Paper's reported PSNR/SSIM on some tasks (e.g., low-light LOL/VE-LOL-L and 25% inpainting SSIM) trail baselines such as Zero-DCE, DDRM or GDP.
- Guidance scale formula is heuristic and derived via Taylor approximation; the Sigma term is dropped without theoretical justification.
- Hardware, runtime and implementation details are not provided.
Experimental setup
Datasets: ImageNet 1k LOL VE-LOL-L LoLi-Phone GoPro HIDE NTIRE LFW WIDER
Benchmarks: LFW WIDER ImageNet 1k LOL VE-LOL-L LoLi-Phone GoPro HIDE NTIRE
Metrics: PSNR SSIM FID NIQE Consistency LOE PI LPIPS
Baselines: PGDiff DiffBIR RED DGP SNIPS DDRM DDNM GDP ExCNet Zero-DCE Zero-DCE++ RRDNet
Builds on: DDPM GDP DDRM DDNM
Closest papers by meaning
- Blind Image Restoration via Fast Diffusion Inversion (NeurIPS 2024 · similarity 0.90)
- A Modular Conditional Diffusion Framework for Image Reconstruction (NeurIPS 2024 · similarity 0.89)
- A Diffusion Model with State Estimation for Degradation-Blind Inverse Imaging (AAAI 2024 · similarity 0.89)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.88)
- Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · similarity 0.88)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.87)
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.