Research map / Image generation / Text-to-image diffusion
EMMA: Concept Erasure Benchmark with Comprehensive Semantic Metrics and Diverse Categories
Lu Wei, Yuta Nakashima, Noa Garcia
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 |
|---|---|---|---|
| use implicit long descriptions instead of the explicit name prompt | MACE · Object | erasing ability: 98.6 -> 70.5 | Hurt |
| use concept remapping (ESD, UCE, MACE) instead of optimization (CA, FMN) | concept erasure methods · Object | erasing ability: 16.5 -> 89.7 | Improved |
| evaluate retention on similar instead of random non-target concepts | MACE · Object | retaining ability: 91.6 -> 79.4 | Hurt |
| evaluate retention on similar instead of random non-target concepts | MACE · Copyright | retaining ability: 36.9 -> 34.7 | Hurt |
| apply FMN to object erasure | FMN · Object | erasing ability: 5.8 | No clear effect |
| use implicit prompts for evaluating CA and FMN | CA and FMN · Object | erasing ability: 16.5 -> 21.2 | No clear effect |
| apply ESD instead of UCE for retention | ESD · Object | retaining ability: 86.0 -> 74.6 | Hurt |
| apply ESD erasure | ESD · bias (gender and ethnicity) | bias | Hurt |
| apply FMN erasure on SD v2.1 | FMN · bias (gender and ethnicity) | bias | Improved |
| apply CE methods | CA, ESD, UCE, MACE · inference time | inference time: 2.5 -> 15.4 | Hurt |
| apply UCE erasure | UCE · Celebrity | erasing ability: 10.3 -> 99.7 | Improved |
| apply MACE for NSFW erasure | MACE · NSFW | erasing ability: 47.1 -> 81.4 | Mixed |
Closest papers by meaning
- Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts (ICCV 2025 · similarity 0.91)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.91)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.91)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- One Image is Worth a Thousand Words: A Usability Preservable Text-Image Collaborative Erasing Framework (ICML 2025 · similarity 0.88)
- Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion models (NeurIPS 2025 · similarity 0.88)
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.
