Research map / Image generation / Text-to-image diffusion
Closing the Safety Gap: Surgical Concept Erasure in Visual Autoregressive Models
Xinhao Zhong, Yimin Zhou, Zhiqi Zhang, Junhao Li, Sun Yi, Bin Chen, Shu-Tao Xia, Xuan Wang, Ke Xu
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 filtered cross-entropy loss L_FCE over the MSE-style baseline (ablation) | S-VARE · Church erasure | ACC_e / FID: ACC_e 9.8 -> 4.1; FID 35.3 -> 32.4 | Improved |
| add preservation loss L_Pre on top of L_FCE (ablation) | S-VARE · Church erasure | ACC_i / FID: ACC_i 73.5 -> 75.7; FID 32.4 -> 31.5 | Improved |
| raise filter ratio alpha from 25% to 75% (ablation) | S-VARE · Ring-A-Bell nudity | ASR: ASR 7.4 -> 40.2 | Hurt |
| set filter ratio alpha to 0% (no filtering) (ablation) | S-VARE · Ring-A-Bell nudity | FID / CLIP: FID 32.8 -> 33.6; CLIP 31.3 -> 30.9 | Hurt |
| feed auxiliary visual tokens from the original model (VARE framework) (ablation) | VAR concept erasure · Infinity-2B qualitative | image quality | Improved |
| fine-tune CA + FFN instead of SA only (ablation) | S-VARE · Ring-A-Bell nudity | ASR: ASR 13.7 -> 7.4 | Improved |
| additionally fine-tune self-attention (SA + CA + FFN) (ablation) | S-VARE · Ring-A-Bell nudity | ASR / FID / CLIP: ASR 7.4 -> 7.1; FID 32.8 -> 33.1; CLIP 31.3 -> 31.0 | Mixed |
| swap losses: KL for erasure and cross-entropy for preservation (ablation) | S-VARE · qualitative | generation quality | Hurt |
| erase 10 Imagenette classes at once instead of one (ablation) | S-VARE · Imagenette | FID: FID 32.4 -> 34.2 | Hurt |
| apply S-VARE to adversarial prompts without adversarial training | S-VARE · Ring-A-Bell | ASR: 75.9 -> 7.4 | Improved |
| use UCE closed-form edit for VAR | UCE baseline · NSFW erasure | Sen.: 158 -> 122 | No clear effect |
Closest papers by meaning
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · 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 0.88)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.88)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.88)
- Consistency-Preserving Concept Erasure via Unsafe–Safe Pairing and Directional Fisher-weighted Adaptation (NeurIPS 2026 · 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.
