Research map / Image generation / Image restoration
Dehazing and deraining
Remove haze and rain using codebook priors, diffusion priors, and unpaired training.
118 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Image restoration, keyword cluster “removal · enhancement · color”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 6 tracked benchmarks.
17 papers in 2023–24 and 41 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 5.5% to 9.4%.
Open questions the data points to
- O-HAZE (PSNR): best went from 24.36 (up to 2024) to 25.72 in 2025 — Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing (AAAI 2025 · 12 citations)
- O-HAZE (SSIM): best went from 0.814 (up to 2024) to 0.88 in 2025 — Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing (AAAI 2025 · 20 citations)
- Papers report RIDCP on RTTS (BRISQUE) anywhere from 15.14 to 18.78: comparisons against it depend on whose number you take.
- Papers report RIDCP on RTTS (FADE) anywhere from 0.48 to 0.944: comparisons against it depend on whose number you take.
What looks settled
- Rain100H (PSNR): best up to 2024 31.51, best in 2025–26 31.79 from 2 papers, less than 1% better.
- Rain100L (PSNR): best up to 2024 40, best in 2025–26 39.73 from 2 papers, at most 0.2% better, if at all.
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
- 2016 · Non-Local Image Dehazing (CVPR 2016 · 1,680 citations)
- 2017 · Removing Rain From Single Images via a Deep Detail Network (CVPR 2017 · 1,298 citations)
- 2018 · Densely Connected Pyramid Dehazing Network (CVPR 2018 · 1,218 citations)
- 2019 · Spatial Attentive Single-Image Deraining With a High Quality Real Rain Dataset (CVPR 2019 · 695 citations)
- 2020 · FFA-Net: Feature Fusion Attention Network for Single Image Dehazing (AAAI 2020 · 1,757 citations)
- 2021 · Contrastive Learning for Compact Single Image Dehazing (CVPR 2021 · 940 citations)
- 2022 · Self-Augmented Unpaired Image Dehazing via Density and Depth Decomposition (CVPR 2022 · 318 citations)
- 2023 · RIDCP: Revitalizing Real Image Dehazing via High-Quality Codebook Priors (CVPR 2023 · 237 citations)
- 2024 · Text-IF: Leveraging Semantic Text Guidance for Degradation-Aware and Interactive Image Fusion (CVPR 2024 · 187 citations)
- 2025 · Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · 40 citations)
- 2026 · SD-PSFNet: Sequential and Dynamic Point Spread Function Network for Image Deraining (AAAI 2026 · 2 citations)
Earliest papers here
- Non-Local Image Dehazing (CVPR 2016 · 1,680 citations)
- Rain Streak Removal Using Layer Priors (CVPR 2016 · 1,002 citations)
- Composition-Preserving Deep Photo Aesthetics Assessment (CVPR 2016 · 291 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.
O-HAZE · PSNR (higher is better)
7 papers, 2019–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 18.82 | "Double-DIP": Unsupervised Image Decomposition via Coupled Deep-Image-Priors (CVPR 2019) |
| 2020 | 24.36 | Multi-Scale Boosted Dehazing Network With Dense Feature Fusion (CVPR 2020 · 1,103 citations) |
| 2025 | 25.72 | Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing (AAAI 2025 · 12 citations) |
Rain100H · PSNR (higher is better)
7 papers, 2020–2025 · small gains only
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 31.28 | A Model-Driven Deep Neural Network for Single Image Rain Removal (CVPR 2020 · 453 citations) |
| 2023 | 31.51 | Hybrid CNN-Transformer Feature Fusion for Single Image Deraining (AAAI 2023 · 79 citations) |
| 2025 | 31.79 | FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining (ICML 2025 · 2 citations) |
Rain100L · PSNR (higher is better)
7 papers, 2020–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 40 | A Model-Driven Deep Neural Network for Single Image Rain Removal (CVPR 2020 · 453 citations) |
Rain200H · PSNR (higher is better)
7 papers, 2020–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 28.16 | Detail-recovery Image Deraining via Context Aggregation Networks (CVPR 2020 · 227 citations) |
| 2021 | 28.87 | Robust Representation Learning With Feedback for Single Image Deraining (CVPR 2021 · 108 citations) |
| 2022 | 29.98 | Online-Updated High-Order Collaborative Networks for Single Image Deraining (AAAI 2022 · 27 citations) |
| 2024 | 32.45 | Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · 47 citations) |
Rain800 · PSNR (higher is better)
7 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 24.59 | Uncertainty Guided Multi-Scale Residual Learning-Using a Cycle Spinning CNN for Single Image De-Raining (CVPR 2019 · 346 citations) |
| 2020 | 26.32 | Detail-recovery Image Deraining via Context Aggregation Networks (CVPR 2020 · 227 citations) |
| 2021 | 27.95 | Robust Representation Learning With Feedback for Single Image Deraining (CVPR 2021 · 108 citations) |
| 2024 | 30.75 | Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · 2 citations) |
O-HAZE · SSIM (higher is better)
6 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 0.749 | Multi-Scale Boosted Dehazing Network With Dense Feature Fusion (CVPR 2020 · 1,103 citations) |
| 2021 | 0.814 | Ultra-High-Definition Image Dehazing via Multi-Guided Bilateral Learning (CVPR 2021 · 260 citations) |
| 2025 | 0.88 | Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing (AAAI 2025 · 20 citations) |
What has been tried
105 papers here report 1,122 measured changes (a component added, removed or swapped, compared with a baseline): helped 80%, no clear effect 2%, hurt 17%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add intensity loss | LLVIP · VIF | 0.794 -> 1.011 | Mixed | Text-IF: Leveraging Semantic Text Guidance for Degradation-Aware and Interactive Image Fusion (CVPR 2024 · 187 citations) |
| add the INR branch | Rain200L · PSNR | 41.40 -> 41.71 | Helped | Bidirectional Multi-Scale Implicit Neural Representations for Image Deraining (CVPR 2024 · 107 citations) |
| Add En_to_DE | UIEB testing set · PSNR | 23.697 -> 23.710 | Helped | Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · 57 citations) |
| Add scale S2 to S1 | Rain200H · PSNR | 32.08 -> 32.20 | Helped | Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · 47 citations) |
| use feature-level rather than image-level depth difference as the prompt | SOTS-indoor · NIQE | 4.901 -> 4.252 | Helped | SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency (AAAI 2024 · 47 citations) |
| Add PAG to the backbone without TAG | Haze2020 · FID | 74.575 -> 73.096 | Helped | Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · 40 citations) |
| Use hybrid conditional and unconditional training | RTTS · LIQE | 2.7754 -> 3.0693 | Helped | Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · 39 citations) |
| use both RGB and YCbCr rather than RGB only | RW²AH · PSNR | 18.44 -> 22.26 | Helped | Guided Real Image Dehazing Using YCbCr Color Space (AAAI 2025 · 38 citations) |
| add Code-Predictor rather than use nearest-neighbor code matching | RTTS · MUSIQ | 58.19 -> 59.60 | Helped | Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing (CVPR 2025 · 36 citations) |
| add model compression (MoC) to the student model | RTTS · FADE | 1.23 -> 1.13 | Helped | CoA: Towards Real Image Dehazing via Compression-and-Adaptation (CVPR 2025 · 28 citations) |
| Add normalizing training (restrict loss and gradient truncation) | DIV2K, realistic conditions, secret/recovery image pair · PSNR | 37.89 -> 53.43 | Helped | StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · 27 citations) |
| Increase Bottleneck depth from 1 to 2 | 8KDehaze · PSNR | 31.61 -> 32.35 | Helped | Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images (CVPR 2025 · 23 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 |
|---|---|---|---|---|
| RIDCP | RTTS | BRISQUE | 15.14 – 18.78 | 3 papers, e.g. AAAI 2025 and AAAI 2025 |
| RIDCP | RTTS | FADE | 0.48 – 0.944 | 2 papers, e.g. AAAI 2025 and AAAI 2025 |
| DAD | RTTS | US | 0.103 – 0.143 | 2 papers, e.g. AAAI 2025 and CVPR 2023 |
| Dehamer | RTTS | US | 0.041 – 0.051 | 2 papers, e.g. CVPR 2023 and AAAI 2025 |
| MSBDN | RTTS | US | 0.037 – 0.046 | 2 papers, e.g. AAAI 2025 and CVPR 2023 |
| PSD | RTTS | US | 0.087 – 0.105 | 2 papers, e.g. AAAI 2025 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 30 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ISTD+ · 4 of 30 papers
- SRD · 4 of 30 papers
- O-HAZE · 3 of 30 papers
- Fattal's dataset · 3 of 30 papers
- RTTS · 3 of 30 papers
- RESIDE · 2 of 30 papers
- NH-HAZE · 2 of 30 papers
- Rain12 · 2 of 30 papers
Baselines
- Dehamer · 9 of 30 papers
- PSD · 8 of 30 papers
- D4 · 8 of 30 papers
- RIDCP · 8 of 30 papers
- MSBDN · 7 of 30 papers
- DAD · 6 of 30 papers
- DehazeFormer · 6 of 30 papers
- ShadowFormer · 5 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 58 papers; from OpenAlex and the papers' affiliations).
- Nanjing University of Science and Technology · 6 papers
- University of Science and Technology of China · 4 papers
- Nanyang Technological University · 4 papers
- Dalian University of Technology · 3 papers
- Nankai University · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Hong Kong University of Science and Technology · 3 papers
- Shenyang Aerospace University · 2 papers
- Wuhan University · 2 papers
- Xiamen University · 2 papers
Use it
Download BibTeX of 40 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
- Inverse imaging with learned priors (109 papers)
- Image harmonization and color transformation (102 papers)
- Reflection removal (95 papers)
- Pansharpening and spectral fusion (84 papers)
- Low-light image enhancement (72 papers)
Elsewhere in the topic
- Efficient restoration architectures (113 papers)
- Image deblurring (73 papers)
- Learned image compression (69 papers)
- Priors for restoration networks (66 papers)
- Diffusion-based universal restoration (61 papers)
- Self-supervised image denoising (60 papers)
- Raw and low-light noise modeling (53 papers)
- Diffusion priors for denoising and deblurring (41 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).