Research map / Image generation / Text-to-image diffusion
Prototype-Guided Concept Erasure in Diffusion Models
Yuze Cai, Jiahao Lu, Hongxiang Shi, Yichao Zhou, Hong Lu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace single-direction erasure with a bank of clustered concept prototypes as negative guidance | SD v1.4 · I2P (overall) | inappropriate rate (%): 35.6% -> 5.2% | Improved |
| apply prototype guidance on SDXL instead of Safree | SDXL · UnlearnDiff attack | ASR: 0.25 -> 0.13 | Improved |
| apply prototype guidance on SD 3.5 instead of Safree | SD 3.5 · P4D attack | ASR: 0.27 -> 0.09 | Improved |
| apply prototype guidance without adversarial-specific design | SD v1.4 · Ring-a-Bell attack | ASR (%): 71.3% -> 6.7% | Improved |
| use prototype guidance instead of TRCE on P4D | SD v1.4 · P4D attack | ASR (%): 2.0% -> 14.5% | Hurt |
| use prototype guidance instead of TRCE on FID | SD v1.4 · adversarial evaluation | FID: 48.7 -> 45.1 | Improved |
| use prototype guidance instead of Safree on FID | SD v1.4 · adversarial evaluation | FID: 36.3 -> 45.1 | Hurt |
| increase number of prototypes from 1 to 16 (ablation) | prototype guidance · I2P sexual | NudeNet rate and CLIP score | Mixed |
| increase number of prototypes beyond 16 (up to 128) (ablation) | prototype guidance · I2P sexual | CLIP score and NudeNet | No clear effect |
| use 16 prototypes for the shocking concept (ablation) | prototype guidance · I2P shocking | Q16 and CLIP score | Mixed |
| use prototype guidance for multi-style erasure | SD v1.4 · Multi-style erasure | Aesthetic score: 5.84 -> 5.89 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Safree — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 4 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.92)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.92)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.91)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.91)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.91)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
