Research map / Architectures and efficiency / Attention and transformers
Efficient image transformers
Applies sparse or shared attention and U-shaped transformers to image restoration and matching.
84 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Attention and transformers, keyword cluster “kv · cache · long context”
State History Benchmarks Tried Baselines Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
36 papers in 2023–24 and 25 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.2% to 2.1%.
Open questions the data points to
- Papers report NetVLAD on Tokyo 24/7 (R@1) anywhere from 64.8 to 95.6: comparisons against it depend on whose number you take.
No longer reported by papers here: DND (PSNR, last in 2024); SIDD (PSNR, last in 2024).
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 · Learning to Match Aerial Images With Deep Attentive Architectures (CVPR 2016 · 87 citations)
- 2021 · Image Super-Resolution With Non-Local Sparse Attention (CVPR 2021 · 611 citations)
- 2022 · Uformer: A General U-Shaped Transformer for Image Restoration (CVPR 2022 · 2,236 citations)
- 2023 · Learning a Sparse Transformer Network for Effective Image Deraining (CVPR 2023 · 525 citations)
- 2024 · Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · 210 citations)
- 2025 · Can We Get Rid of Handcrafted Feature Extractors? SparseViT: Nonsemantics-Centered, Parameter-Efficient Image Manipulation Localization Through Spare-Coding Transformer (AAAI 2025 · 40 citations)
- 2026 · Collaborative Transformers with Multi-Level Forensic Attention for Image Manipulation Localization (AAAI 2026 · 1 citation)
Earliest papers here
- Learning to Match Aerial Images With Deep Attentive Architectures (CVPR 2016 · 87 citations)
- Image Super-Resolution With Non-Local Sparse Attention (CVPR 2021 · 611 citations)
- Pose Recognition With Cascade Transformers (CVPR 2021 · 293 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.
DND · PSNR (higher is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 40.05 | Cross Aggregation Transformer for Image Restoration (NeurIPS 2022 · 122 citations) |
SIDD · PSNR (higher is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 40.01 | Cross Aggregation Transformer for Image Restoration (NeurIPS 2022 · 122 citations) |
What has been tried
70 papers here report 774 measured changes (a component added, removed or swapped, compared with a baseline): helped 78%, no clear effect 4%, hurt 15%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| replace Swin self-attention with ASSA | SPAD · PSNR | 44.47 -> 45.43 | Helped | Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · 210 citations) |
| remove dual-softmax at inference | MegaDepth · Pose estimation AUC@5° | 56.4 -> 55.4 | Hurt | Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed (CVPR 2024 · 177 citations) |
| Add image-level and step-level correlation information | Mindboggle, Buckner, and ACDC · DSC | 0.628 -> 0.642; 0.650 -> 0.661; 0.800 -> 0.810 | Helped | Correlation-aware Coarse-to-fine MLPs for Deformable Medical Image Registration (CVPR 2024 · 62 citations) |
| Increase the number of learnable queries from 8 to 64 per BoQ block | MSLS-val · Recall@1 | 88.1 -> 91.3 | Helped | BoQ: A Place is Worth a Bag of Learnable Queries (CVPR 2024 · 42 citations) |
| replace global attention with sparse attention | COVERAGE · pixel-level F1 | 0.378 -> 0.485 | Helped | Can We Get Rid of Handcrafted Feature Extractors? SparseViT: Nonsemantics-Centered, Parameter-Efficient Image Manipulation Localization Through Spare-Coding Transformer (AAAI 2025 · 40 citations) |
| Add random shuffle and inverse shuffle to local self-attention | ISTD+ · MAE | 6.4 -> 5.0 (shadow); 2.5 -> 2.3 (non-shadow); 3.0 -> 2.7 (all) | Helped | HomoFormer: Homogenized Transformer for Image Shadow Removal (CVPR 2024 · 39 citations) |
| Add dilated M-WSA to the dilated-convolution base | SIDD validation · PSNR | +0.27dB | Helped | Rethinking Transformer-Based Blind-Spot Network for Self-Supervised Image Denoising (AAAI 2025 · 32 citations) |
| Add AIRB | Set5, Set11, and McM18, δ=0.25 · PSNR | 36.88 -> 37.08; 35.16 -> 35.42; 37.03 -> 37.24 | Helped | UFC-Net: Unrolling Fixed-point Continuous Network for Deep Compressive Sensing (CVPR 2024 · 28 citations) |
| add the occlusion-embedded mix layer | BasicLFSR ×4 SR · PSNR (dB) | 32.30 -> 32.55 | Helped | Occlusion-Embedded Hybrid Transformer for Light Field Super-Resolution (AAAI 2025 · 27 citations) |
| Add L/Swin-FA | KAIST 10-scene simulation, Avg · PSNR | 37.28 -> 38.49 | Helped | Dual Prior Unfolding for Snapshot Compressive Imaging (CVPR 2024 · 20 citations) |
| add the fine-pixel-refinement enhancement branch | KAIST simulation · PSNR (dB) | 32.23 -> 32.34 | Helped | Dual-Window Multiscale Transformer for Hyperspectral Snapshot Compressive Imaging (AAAI 2024 · 15 citations) |
| add cross-stage skip connections (CSSC) | Set11, 10% SR · PSNR | 31.48 -> 32.26; 0.78 dB | Helped | Dual-Scale Transformer for Large-Scale Single-Pixel Imaging (CVPR 2024 · 15 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 |
|---|---|---|---|---|
| NetVLAD | Tokyo 24/7 | R@1 | 64.8 – 95.6 | 2 papers, e.g. CVPR 2023 and NeurIPS 2025 |
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 61 papers; from OpenAlex and the papers' affiliations).
- Harbin Institute of Technology · 6 papers
- Westlake University · 4 papers
- Shanghai Jiao Tong University · 4 papers
- University of Science and Technology of China · 4 papers
- Zhejiang University · 4 papers
- Microsoft · 4 papers
- The University of Sydney · 3 papers
- Peng Cheng Laboratory · 3 papers
- Northwestern Polytechnical University · 3 papers
- ByteDance · 2 papers
Use it
Download BibTeX of 38 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
- Sparse and structured attention (163 papers)
- Attention head and block redesign (127 papers)
- Long-context KV and memory attention (88 papers)
- Transformer neural operators for PDEs (69 papers)
Elsewhere in the topic
- Transformer expressivity and learnability (217 papers)
- State space and recurrent sequence models (172 papers)
- Sparse attention for long-context LLMs (147 papers)
- Efficient and robust vision transformers (124 papers)
- Efficient language transformers (105 papers)
- Language model pretraining and induction (99 papers)
- Hybrid and interpretable ViT designs (94 papers)
- KV cache eviction and compression (92 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).