Research map / Image generation / Editing and style transfer
Inversion-based diffusion editing
Edit real images through diffusion inversion and training-free manipulation.
113 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Editing and style transfer, keyword cluster “image editing · edits · inversion”
State History Tried Baselines Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
45 papers in 2023–24 and 66 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 16.5% to 16.8%.
Open questions the data points to
- Papers report DragDiffusion on DragBench (Time (s)) anywhere from 32.93 to 69.4: comparisons against it depend on whose number you take.
- Papers report SDS on SVG editing (User Preference Rate) anywhere from 26.51 to 30.83: 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
- 2022 · DiffusionCLIP: Text-Guided Diffusion Models for Robust Image Manipulation (CVPR 2022 · 514 citations)
- 2023 · RODIN: A Generative Model for Sculpting 3D Digital Avatars Using Diffusion (CVPR 2023 · 205 citations)
- 2024 · DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · 145 citations)
- 2025 · DiT4Edit: Diffusion Transformer for Image Editing (AAAI 2025 · 36 citations)
- 2026 · TweezeEdit: Consistent and Efficient Image Editing with Path Regularization (AAAI 2026 · 2 citations)
Earliest papers here
- DiffusionCLIP: Text-Guided Diffusion Models for Robust Image Manipulation (CVPR 2022 · 514 citations)
- Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models (NeurIPS 2022 · 17 citations)
- RODIN: A Generative Model for Sculpting 3D Digital Avatars Using Diffusion (CVPR 2023 · 205 citations)
What has been tried
82 papers here report 848 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 5%, hurt 22%, 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 |
|---|---|---|---|---|
| add identity-preserving fine-tuning | DRAGBENCH · Mean Distance | Helped | DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing (CVPR 2024 · 145 citations) | |
| Use the image prompt rather than text prompting for DDIM inversion and editing | Image reconstruction and fine-grained editing · reconstruction/editing quality | Helped | DiffEditor: Boosting Accuracy and Flexibility on Diffusion-based Image Editing (CVPR 2024 · 44 citations) | |
| Add patches merging | 512×512, 1024×1024, and 1024×2048 · Speed | 9.13 -> 6.01; 41.52 -> 34.27; 122.41 -> 99.39 | Helped | DiT4Edit: Diffusion Transformer for Image Editing (AAAI 2025 · 36 citations) |
| add LoRA interpolation to DDIM | MorphBench Overall · FID | 101.68 -> 44.40 | Helped | DiffMorpher: Unleashing the Capability of Diffusion Models for Image Morphing (CVPR 2024 · 33 citations) |
| Replace a single encoder with separate map-specific encoders | Material map reconstruction · RMSE (diffuse) | 0.047 -> 0.016 | Helped | MatFuse: Controllable Material Generation with Diffusion Models (CVPR 2024 · 33 citations) |
| Replace the DDIM inversion prior with random initialization of z_T | Object moving · content consistency | Hurt | DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models (ICLR 2024 · 33 citations) | |
| Set regularization ratio λ to 0.1 | MS-COCO val · ISTD | 16.54 -> 24.23 | Hurt | Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · 30 citations) |
| Add self-attention CUT loss to vanilla DDS | image editing · structural consistency | Helped | Contrastive Denoising Score for Text-guided Latent Diffusion Image Editing (CVPR 2024 · 29 citations) | |
| Use local foreground guidance instead of average foreground guidance | Ablation Study · identity preservation | Helped | Diffusion Handles Enabling 3D Edits for Diffusion Models by Lifting Activations to 3D (CVPR 2024 · 25 citations) | |
| Use vital-layer attention injection rather than injecting in all layers | COCO evaluation dataset · CLIPtxt | 0.17 -> 0.23 | Helped | Stable Flow: Vital Layers for Training-Free Image Editing (CVPR 2025 · 21 citations) |
| Replace P2P with UAC | PIE-Bench · PSNR (10^3) | 26.64 -> 28.51 | Helped | Inversion-Free Image Editing with Language-Guided Diffusion Models (CVPR 2024 · 21 citations) |
| start the editing stage at different initial timesteps | point-based image editing · scale of controlled semantics | Mixed | Drag Your Noise: Interactive Point-based Editing via Diffusion Semantic Propagation (CVPR 2024 · 16 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 |
|---|---|---|---|---|
| DragDiffusion | DragBench | Time (s) | 32.93 – 69.4 | 2 papers, e.g. ICML 2025 and CVPR 2024 |
| SDS | SVG editing | User Preference Rate | 26.51 – 30.83 | 2 papers, e.g. ICML 2025 and CVPR 2024 |
| DDS | SVG editing | LPIPS | 0.524 – 0.598 | 2 papers, e.g. ICML 2025 and CVPR 2024 |
| SDS | SVG editing | LPIPS | 0.438 – 0.485 | 2 papers, e.g. ICML 2025 and CVPR 2024 |
What reviewers will expect
Datasets and baselines named in the 39 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- PIE-Bench · 6 of 39 papers
- CelebA-HQ · 5 of 39 papers
- COCO · 2 of 39 papers
- TEdBench · 2 of 39 papers
- LSUN-church · 2 of 39 papers
- LSUN-bedroom · 2 of 39 papers
- InstructPix2Pix dataset · 2 of 39 papers
- DreamBooth · 2 of 39 papers
Baselines
- SDEdit · 7 of 39 papers
- AnyDoor · 6 of 39 papers
- Paint-by-Example · 5 of 39 papers
- MasaCtrl · 5 of 39 papers
- DragDiffusion · 4 of 39 papers
- DragonDiffusion · 4 of 39 papers
- DiffEditor · 4 of 39 papers
- DDIM inversion · 4 of 39 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 111 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 5 papers
- Tencent · 5 papers
- Adobe · 5 papers
- Hong Kong University of Science and Technology · 4 papers
- Peking University · 3 papers
- Zhejiang University · 3 papers
- Chinese Academy of Sciences · 3 papers
- University of Michigan · 3 papers
- Korea University · 3 papers
- The Chinese University of Hong Kong · 3 papers
Use it
Download BibTeX of 27 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
- Instruction-based diffusion editing (151 papers)
- Localized and attribute-specific editing (73 papers)
- GAN inversion for editing (25 papers)
Elsewhere in the topic
- Texture and sketch synthesis (130 papers)
- Image-to-image translation (127 papers)
- Arbitrary neural style transfer (124 papers)
- Identity-preserving face manipulation (88 papers)
- Diffusion face generation and editing (61 papers)
- Virtual try-on (46 papers)
- Pose-guided person image synthesis (29 papers)
- Talking face generation (21 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).