Research map / Image generation / Editing and style transfer
Orthogonal Concept Erasure for Diffusion Models
Yuhao Sun, Lingyun Yu, Hao-Xiang Xu, Fengyuan Miao, Zhuoer Xu, Hongtao Xie
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 additive parameter update with layer-wise orthogonal update (OCE) | editing-based concept erasure · CIFAR-10 object erasure (avg of 5 classes) | Acc_e: 8.32 (MACE) -> 4.61 | Improved |
| use subspace-level erasure objective with vector-wise preservation instead of vector-wise erasure and preservation (ablation) | OCE · Erase 100 celebrities | Acc_e: 7.59 -> 3.44 | Improved |
| use subspace-level erasure with vector-wise preservation rather than subspace-level preservation (ablation) | OCE · Erase 100 celebrities | Ho: 94.22 -> 95.48 | Improved |
| subspace-level preservation instead of vector-wise preservation (ablation) | OCE · Erase 100 celebrities | FID: 18.10 vs 18.33 FID; Acc_e 4.54 vs 3.44 | Mixed |
| add global preservation prior K0 from COCO-30K tokens (ablation) | OCE · Erase 100 celebrities | Ho: 93.80 -> 95.48 | Improved |
| add global preservation prior K0 (ablation) | OCE · Erase 100 celebrities (MSCOCO-30k) | FID: 22.76 -> 18.33 | Improved |
| increase generic concept set for K0 from 1/3 to full (ablation) | OCE · Erase 100 celebrities | Acc_e: 4.47 -> 3.44 | Improved |
| add adversarial editing (AT) to OCE (ablation) | OCE · I2P implicit concept erasure | ASR: 0.11 -> 0.05 | Improved |
| add adversarial editing (AT) to OCE (ablation) | OCE · MSCOCO-1K preservation | CLIP score: 26.33 -> 26.10 | Hurt |
| use OCE instead of MACE for 100-celebrity erasure | celebrity erasure · Erase 100 celebrities | Time (s): 1800 -> 4.3 | Improved |
| rotate neurons with a shared layer-wise orthogonal matrix vs neuron-wise rotation (toy experiment) (ablation) | SD cross-attention key/value matrices · cat concept generation (qualitative) | image quality | Hurt |
| scale weights by alpha=0.5 (magnitude only) (ablation) | SD cross-attention key/value matrices · cat concept generation (qualitative) | concept semantics | No clear effect |
Closest papers by meaning
- When Are Concepts Erased From Diffusion Models? (NeurIPS 2025 · similarity 0.91)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.90)
- Fantastic Targets for Concept Erasure in Diffusion Models and Where To Find Them (ICLR 2025 · similarity 0.90)
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.90)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 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.
