Research map / Image generation / Image restoration
Diffusion-based universal restoration
Apply generative diffusion priors to unified blind image restoration.
61 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Image restoration, keyword cluster “image restoration · degradation · degraded”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 3 tracked benchmarks.
25 papers in 2023–24 and 34 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 8.0% to 7.8%.
Open questions the data points to
- WIDER-Test (FID): best went from 36.91 (up to 2024) to 35.43 in 2025 — $InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration (ICLR 2025 · 0 citations)
- WebPhoto-Test (FID): best went from 81.53 (up to 2024) to 75.48 in 2025 — $InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration (ICLR 2025 · 0 citations)
- Papers report DifFace on LFW-Test (FID) anywhere from 45.23 to 67.98: comparisons against it depend on whose number you take.
- Papers report CodeFormer on WebPhoto-Test (MUSIQ) anywhere from 55.2 to 74: comparisons against it depend on whose number you take.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Deep Mean-Shift Priors for Image Restoration (NeurIPS 2017 · 67 citations)
- 2022 · Denoising Diffusion Restoration Models (NeurIPS 2022 · 234 citations)
- 2023 · Generative Diffusion Prior for Unified Image Restoration and Enhancement (CVPR 2023 · 275 citations)
- 2024 · Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · 187 citations)
- 2025 · UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior (CVPR 2025 · 18 citations)
- 2026 · MAISI-v2: Accelerated 3D High-Resolution Medical Image Synthesis with Rectified Flow and Region-specific Contrastive Loss (AAAI 2026 · 9 citations)
Earliest papers here
- Deep Mean-Shift Priors for Image Restoration (NeurIPS 2017 · 67 citations)
- Denoising Diffusion Restoration Models (NeurIPS 2022 · 234 citations)
- Generative Diffusion Prior for Unified Image Restoration and Enhancement (CVPR 2023 · 275 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
LFW-Test · FID (lower is better)
5 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 50.98 | DR2: Diffusion-Based Robust Degradation Remover for Blind Face Restoration (CVPR 2023 · 93 citations) |
| 2024 | 43.17 | WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · 10 citations) |
WIDER-Test · FID (lower is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 39.17 | PGDiff: Guiding Diffusion Models for Versatile Face Restoration via Partial Guidance (NeurIPS 2023) |
| 2024 | 36.91 | WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · 10 citations) |
| 2025 | 35.43 | $InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration (ICLR 2025 · 0 citations) |
WebPhoto-Test · FID (lower is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 86.18 | PGDiff: Guiding Diffusion Models for Versatile Face Restoration via Partial Guidance (NeurIPS 2023) |
| 2024 | 81.53 | WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · 10 citations) |
| 2025 | 75.48 | $InterLCM$: Low-Quality Images as Intermediate States of Latent Consistency Models for Effective Blind Face Restoration (ICLR 2025 · 0 citations) |
What has been tried
37 papers here report 382 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 4%, hurt 21%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add self-attention to WTB | UIEBD · PSNR | 20.94 -> 21.87 | Helped | Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · 187 citations) |
| predict noise only instead of residuals only | CelebA 64×64 · FID | 31.47 -> 23.25 | Helped | Residual Denoising Diffusion Models (CVPR 2024 · 138 citations) |
| Add the data fidelity term and TV regularization to the guidance function | Denoising, super-resolution, inpainting · PSNR | 10.12 -> 36.14; 11.49 -> 30.70; 11.13 -> 31.71 | Helped | HIR-Diff: Unsupervised Hyperspectral Image Restoration Via Improved Diffusion Models (CVPR 2024 · 53 citations) |
| remove Stage I training | Outdoor-Rain · PSNR | 30.99 -> 28.90 | Hurt | Learning Diffusion Texture Priors for Image Restoration (CVPR 2024 · 33 citations) |
| Add CFRM feature restoration | DIV2K test · PSNR | 21.43 -> 24.32 | Helped | UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior (CVPR 2025 · 18 citations) |
| Set diffusion step size δt to 50 instead of 100 | CelebA, Denoising · PSNR | 28.67 -> 28.74 | No clear effect | Blind Image Restoration via Fast Diffusion Inversion (NeurIPS 2024 · 17 citations) |
| Add a generative album close to the input to constrain the generative space | Single-image restoration · image quality and input fidelity | Helped | Restoration by Generation with Constrained Priors (CVPR 2024 · 12 citations) | |
| replace text-only prompting with visual-text dual prompts | DRealSR synthetic · LPIPS | 0.4217 -> 0.3846 | Helped | Dual Prompting Image Restoration with Diffusion Transformers (CVPR 2025 · 11 citations) |
| Remove RLDM | LOL-v2-real · PSNR | 21.25 -> 22.97 | Hurt | Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model (ICLR 2025 · 10 citations) |
| Use DWT level 1 instead of level 2 | CelebA-Test · PSNR | 26.97 -> 27.08 | Helped | WaveFace: Authentic Face Restoration with Efficient Frequency Recovery (CVPR 2024 · 10 citations) |
| Use stochastic sampling instead of deterministic regression | FFHQ 256×256, super-resolution (8×) · PSNR | 22.22 -> 21.08 | Hurt | Reconciling Stochastic and Deterministic Strategies for Zero-shot Image Restoration using Diffusion Model in Dual (CVPR 2025 · 6 citations) |
| Use Real-ESRNet output as the initial prediction rather than x0 initialization at T′=T=50 | ImageNet validation, ×4 noisy SR (σn=0.03) · PSNR | 22.38 -> 24.68 | Helped | Improving Diffusion-Based Image Restoration with Error Contraction and Error Correction (AAAI 2024 · 6 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| DifFace | LFW-Test | FID | 45.23 – 67.98 | 2 papers, e.g. CVPR 2024 and NeurIPS 2023 |
| CodeFormer | WebPhoto-Test | MUSIQ | 55.2 – 74 | 2 papers, e.g. CVPR 2024 and ICLR 2025 |
| CodeFormer | WebPhoto-Test | FID | 86.19 – 98.55 | 2 papers, e.g. NeurIPS 2023 and CVPR 2024 |
What reviewers will expect
Datasets and baselines named in the 15 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- FFHQ · 4 of 15 papers
- CelebA-HQ · 3 of 15 papers
- DIV2K · 2 of 15 papers
- CelebA · 2 of 15 papers
- ImageNet · 2 of 15 papers
- fastMRI knee · 2 of 15 papers
Baselines
- DPS · 5 of 15 papers
- DiffBIR · 4 of 15 papers
- DiffPIR · 4 of 15 papers
- DDNM · 3 of 15 papers
- DDRM · 3 of 15 papers
- RePaint · 2 of 15 papers
- DiffIR · 2 of 15 papers
- GDP · 2 of 15 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 59 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Tsinghua University · 2 papers
- Shanghai Jiao Tong University · 2 papers
- University of Texas at Austin · 2 papers
- Fudan University · 2 papers
- Xi’an Jiaotong University · 2 papers
- Huawei · 2 papers
- NVIDIA · 2 papers
- Microsoft · 2 papers
Use it
Download BibTeX of 29 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Efficient restoration architectures (113 papers)
- Priors for restoration networks (66 papers)
- Real-world restoration with unaligned data (40 papers)
Elsewhere in the topic
- Dehazing and deraining (118 papers)
- Inverse imaging with learned priors (109 papers)
- Image harmonization and color transformation (102 papers)
- Reflection removal (95 papers)
- Pansharpening and spectral fusion (84 papers)
- Image deblurring (73 papers)
- Low-light image enhancement (72 papers)
- Learned image compression (69 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).