Research map / Image generation / Autoregressive and latent generation
ScaleErasure: Inference-Time Minimal Intervention for Precise Concept Erasure in Next-Scale Autoregressive Image Generation
Cong Wang, Haiyu Wu, Zhiwei Jiang, Zifeng Cheng, Fei Shen, Yafeng Yin, Qing Gu
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 token-level logits selection (ablation) | ScaleErasure · MS-COCO | FID: 73.11 -> 2.91 | Improved |
| add token-level logits selection (ablation) | ScaleErasure · I2P | ASR (%): 3.65% -> 9.99% | Hurt |
| add bit-channel-level logits selection (ablation) | ScaleErasure · MS-COCO | FID: 3.08 -> 2.91 | Improved |
| add bit-channel-level logits selection (ablation) | ScaleErasure · I2P | ASR (%): 8.49% -> 9.99% | Hurt |
| add penalization in safe guidance (ablation) | ScaleErasure · I2P | ASR (%): 22.56% -> 9.99% | Improved |
| add penalization in safe guidance (ablation) | ScaleErasure · MS-COCO | FID: 2.94 -> 2.91 | No clear effect |
| add safe logits substitution in safe guidance (ablation) | ScaleErasure · I2P | ASR (%): 20.52% -> 9.99% | Improved |
| add safe logits substitution in safe guidance (ablation) | ScaleErasure · MS-COCO | FID: 2.34 -> 2.91 | Hurt |
| increase penalization weight lambda from 0.0 to 5.0 (ablation) | ScaleErasure · I2P | ASR (%): 22.56 -> 8.92 | Improved |
| increase penalization weight lambda from 0.0 to 5.0 (ablation) | ScaleErasure · MS-COCO | FID: 52.92 -> 52.39 | No clear effect |
| use ScaleErasure instead of SLD-strong | ScaleErasure · I2P nudity | NudeNet total nude count: 474 -> 182 | Improved |
| use ScaleErasure instead of ESD-u | ScaleErasure · MS-COCO | FID: 6.41 -> 2.91 | Improved |
Closest papers by meaning
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.90)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.87)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.87)
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.87)
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.87)
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.
