Research map / Image generation / Text-to-image diffusion
SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models
Kien Nguyen, Anh Tran, Cuong Pham
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 reference subspace from TI step 50 instead of step 70 (ablation) | SuMa · Subclass erasure | ASR: 0.56 -> 0.12 | Improved |
| use reference subspace from TI step 50 instead of step 30 (ablation) | SuMa · Subclass erasure | FID: 19.11 -> 18.33 | Improved |
| use TI early-step token instead of general term as reference (ablation) | SuMa · Subclass erasure | ASR: 0.18 -> 0.12 | Improved |
| use TI early-step token instead of general term as reference (ablation) | SuMa · Identity erasure | ASR: 0.17 -> 0.09 | Improved |
| add subspace regularization (ablation) | SuMa · Subclass erasure | FID: 19.21 -> 18.33 | Improved |
| add subspace regularization (ablation) | SuMa · Identity erasure | FID: 20.31 -> 17.94 | Improved |
| add subspace regularization (ablation) | SuMa · Instance erasure | FID: 24.57 -> 22.34 | Improved |
| add subspace regularization (ablation) | SuMa · Instance erasure | CCE ASR: 0.15 -> 0.17 | Hurt |
| use CA loss as base method instead of UCE (ablation) | SuMa · Identity erasure (Elon Musk) | CCE ASR: 0.49 -> 0.07 | Improved |
| use CA loss as base method instead of ESD (ablation) | SuMa · Instance erasure (Grumpy Cat) | FID: 24.57 -> 22.34 | Improved |
| apply L_sub alone without L_CA (ablation) | SuMa · Identity erasure (Elon Musk) | ASR: 0.09 -> 0.73 | Hurt |
| apply L_sub alone without L_CA (ablation) | SuMa · Artistic erasure (Van Gogh) | ASR: 0.05 -> 0.76 | Hurt |
Later papers that compared with this method
1 later accepted paper reports SUMA as a baseline in their results tables; 0 of them report a better number on most of the results they share with it.
- Push–Pull Attentional Anchoring for Diffusion Concept Erasure (ECCV 2026 · better on 1 of 3 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
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.91)
- Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion models (NeurIPS 2025 · similarity 0.90)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.90)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.89)
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.
