Research map / Image generation / Text-to-image diffusion
Mass Concept Erasure in Diffusion Models with Concept Hierarchy
Jiahang Tu, Ye Li, Yiming Wu, Hanbin Zhao, Chao Zhang, Hui Qian
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 |
|---|---|---|---|
| add concept hierarchy with group-wise erasure (ablation) | SuPLoRA erasure · Mass erasure benchmark | parameters (MB): 28.5MB -> 7.11MB | Improved |
| add concept hierarchy with group-wise erasure (ablation) | SuPLoRA erasure · Mass erasure benchmark | erasure accuracy: 6.040% -> 5.830% | No clear effect |
| add SuPLoRA (ablation) | hierarchical group-wise erasure · Mass erasure benchmark | domain-specific accuracy: 67.81% -> 72.94% | Improved |
| add SuPLoRA (ablation) | hierarchical group-wise erasure · Mass erasure benchmark | supertype CLIP score: 25.19 -> 26.21 | Improved |
| add diffusion regularization loss (ablation) | SuPLoRA erasure · Mass erasure benchmark | MS-COCO FID: 18.97 -> 17.94 | Improved |
| add diffusion regularization loss (ablation) | SuPLoRA erasure · Mass erasure benchmark | erasure accuracy: 4.360% -> 5.830% | Hurt |
| train both A and B instead of freezing B (ablation) | SuPLoRA · Mass erasure benchmark | supertype CLIP score: 26.21 -> 25.22 | Hurt |
| freeze random down-projection B (ablation) | standard LoRA · Mass erasure benchmark | supertype CLIP score: 25.19 -> 26.08 | Improved |
| initialize frozen B with orthogonal complement of supertype subspace (ablation) | LoRA with frozen B · Mass erasure benchmark | supertype CLIP score: 26.08 -> 26.21 | Improved |
| use SuPLoRA instead of MACE | mass erasure · Mass erasure benchmark | retained celebrity accuracy: 78.50% -> 83.38% | Improved |
| use SuPLoRA instead of MACE | mass erasure · Mass erasure benchmark | celebrity erasure accuracy: 6.250% -> 7.500% | Hurt |
| use SuPLoRA instead of MACE | mass erasure · Mass erasure benchmark | storage (MB): 198 -> 154 | Improved |
Closest papers by meaning
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.92)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.90)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (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 0.90)
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.
