Research map / Image generation / Text-to-image diffusion
Concept erasure in diffusion models
Remove unwanted concepts from diffusion models via editing, adapters, and unlearning.
76 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · an approach in Text-to-image diffusion, keyword cluster “concept · erasure · unlearning”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 1 tracked benchmark.
13 papers in 2023–24 and 63 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.3% to 7.4%.
Open questions the data points to
- COCO-30k (FID): best went from 15.98 (up to 2024) to 1.44 in 2025 — CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · 1 citation)
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
- 2024 · MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · 73 citations)
- 2025 · Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · 14 citations)
- 2026 · TokenTrace: Multi-Concept Attribution through Watermarked Token Recovery (CVPR 2026 · 1 citation)
Earliest papers here
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · 73 citations)
- Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation (CVPR 2024 · 30 citations)
- One-dimensional Adapter to Rule Them All: Concepts Diffusion Models and Erasing Applications (CVPR 2024 · 28 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.
COCO-30k · FID (lower is better)
6 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 15.98 | Self-Discovering Interpretable Diffusion Latent Directions for Responsible Text-to-Image Generation (CVPR 2024 · 30 citations) |
| 2025 | 1.44 | CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · 1 citation) |
What has been tried
39 papers here report 384 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 5%, hurt 24%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Remove LoRA training, fusion, and CFIS, leaving closed-form refinement only | 100 celebrities · Hc (%) | 46.72 -> 89.78 | Hurt | MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · 73 citations) |
| add utility-retaining regularization with COCO prompts | Nudity unlearning · FID | 26.48 -> 19.88 | Helped | Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · 19 citations) |
| Increase the unconditional-score regularization coefficient to λ=1.5 | Nudity erasure · spatial consistency | Helped | All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · 16 citations) | |
| Add concept-preserving gradient rather than omit retaining operations | Snoopy unlearning · FID | 146.53 -> 204.37 | Helped | Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · 14 citations) |
| Disable the text filter during prompt optimization, then restore it for evaluation | I2P nudity (SLD-MAX, SLD-STRONG, SD-NEGP) · failure rate | 25.67% -> 40.98%; 34.03% -> 50.25%; 25.44% -> 27.93% (P4D-N); 38.69% -> 39.11%; 37.84% -> 42.79%; 20.36% -> 32.46% (P4D- | Helped | Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · 8 citations) |
| Add adjacency loss (Ladj) | Stanford Dogs — Welsh Springer Spaniel · ERB | 38.58 -> 95.97 | Helped | Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · 6 citations) |
| Map Golf Ball to the semantically distant Outdoor Recreational Activities instead of the closer Tennis Ball | Golf Ball · revival point | 1 -> R | Helped | The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · 5 citations) |
| add target principal components Vtar | Localized celebrities: 50 Celebrities Erased · target ACC | 2.17 -> 1.17 | Helped | Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · 4 citations) |
| Increase prompt length from K=16 to K=77 | Nudity, ESD (w/o SC) · ASR | 35.79% -> 20% | Hurt | Ring-A-Bell! How Reliable are Concept Removal Methods For Diffusion Models? (ICLR 2024 · 4 citations) |
| add diverse remaining data | CIFAR-10 class-wise forgetting · FID | 9.813 (±1.863) -> 6.539 (±0.994) | Helped | Boosting Alignment for Post-Unlearning Text-to-Image Generative Models (NeurIPS 2024 · 3 citations) |
| Add output-preserving regularization | MS COCO generation during nudity unlearning · prior preservation | Helped | Direct Unlearning Optimization for Robust and Safe Text-to-Image Models (NeurIPS 2024 · 3 citations) | |
| Add timestep modulation | Van Gogh · LPIPS_da | 0.184 -> 0.318 | Helped | DuMo: Dual Encoder Modulation Network for Precise Concept Erasure (AAAI 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 32 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- I2P · 20 of 32 papers
- COCO-30K · 8 of 32 papers
- Imagenette · 5 of 32 papers
- Ring-A-Bell · 4 of 32 papers
- MMA-Diffusion · 4 of 32 papers
- COCO-10K · 3 of 32 papers
- COCO · 2 of 32 papers
- ImageNet-1k · 2 of 32 papers
Baselines
- ESD · 20 of 32 papers
- UCE · 20 of 32 papers
- MACE · 15 of 32 papers
- SPM · 14 of 32 papers
- SLD · 10 of 32 papers
- RECE · 8 of 32 papers
- CA · 7 of 32 papers
- ESD-u · 7 of 32 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 76 papers; from OpenAlex and the papers' affiliations).
- Nanyang Technological University · 4 papers
- Fudan University · 3 papers
- University of California, Berkeley · 3 papers
- Alibaba Group · 3 papers
- Michigan State University · 3 papers
- New York University · 3 papers
- IBM · 3 papers
- Tsinghua University · 2 papers
- A*STAR · 2 papers
- University of Oxford · 2 papers
Use it
Download BibTeX of 20 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
- Multi-concept composition (56 papers)
Elsewhere in the topic
- Compositional text-to-image guidance (194 papers)
- Layout-to-image control (164 papers)
- T2I evaluation and controllable adapters (149 papers)
- Text-driven synthesis and semantic guidance (137 papers)
- Personalized and region-controlled generation (136 papers)
- Inference-time T2I improvement (117 papers)
- Multi-subject personalization (82 papers)
- Preference optimization for diffusion (82 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).