Research map / Image generation / Text-to-image diffusion
Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts
Ibtihel Amara, Ahmed Imtiaz Humayun, Ivana Kajic, Zarana Parekh, Natalie Harris, Sarah Young, Chirag Nagpal, Najoung Kim, Junfeng He, Cristina Nader Vasconcelos, Deepak Ramachandran, Golnoosh Farnadi and 3 more
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 |
|---|---|---|---|
| erase a concept with UCE (sensitivity vs original SD) | UCE · EraseBench Visual Similarity (Object) | CLIP sensitivity: 85.0 -> 82.7 | Hurt |
| erase a concept with Receler (sensitivity vs original SD) | Receler · EraseBench Subset-Superset | CLIP sensitivity: 92.1 -> 36.7 | Hurt |
| erase a concept with AdvUnlearn (sensitivity vs original SD) | AdvUnlearn · EraseBench Artistic Style Similarity | CLIP sensitivity: 69.0 -> 27.4 | Hurt |
| erase a concept with ESD (sensitivity vs original SD) | ESD · EraseBench Binomial | CLIP sensitivity: 88.5 -> 70.6 | Hurt |
| erase concepts with UCE, Gecko score on non-erased prompts | UCE · EraseBench Gecko | Gecko alignment: 77.6 -> 74.3 | Hurt |
| erase concepts with MACE, Gecko score on non-erased prompts | MACE · EraseBench Gecko | Gecko alignment: 77.6 -> 67.9 | Hurt |
| erase concepts with AdvUnlearn, Gecko score on non-erased prompts | AdvUnlearn · EraseBench Gecko | Gecko alignment: 77.6 -> 68.6 | Hurt |
| erase concepts with MACE, object presence question accuracy on visually similar prompts | MACE · EraseBench Gecko object presence | Gecko object-presence accuracy: 89.0% -> 67.1% | Hurt |
| erase concepts with UCE, object presence question accuracy on visually similar prompts | UCE · EraseBench Gecko object presence | Gecko object-presence accuracy: 89.0% -> 80.3% | Hurt |
| erase 'cat' with UCE, RAHF artifact score | UCE · EraseBench Erase Cat | RAHF artifacts: 87.71 -> 72.44 | Hurt |
| erase 'goldfish' with AdvUnlearn, RAHF aesthetics | AdvUnlearn · EraseBench Erase Goldfish | RAHF aesthetics: 78.21 -> 72.49 | Hurt |
| add related non-target concepts (tiger, panther) to the retain set (ablation) | erasure with retain set · EraseBench erase cat | over-erasure and artifacts: partially alleviates over-erasure, artifacts persist, erased concept leaks back | Mixed |
Closest papers by meaning
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.91)
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.91)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.91)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.91)
- One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework (ICML 2025 · similarity 0.91)
- When Are Concepts Erased From Diffusion Models? (NeurIPS 2025 · similarity 0.90)
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.
