Research map / Image generation / Text-to-image diffusion
Implicit Concept Removal of Diffusion Models
Zhili Liu, Kai Chen, Yifan Zhang, Jianhua Han, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James Kwok
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 geometric location tokens and concept name to captions, used as negative prompt | Stable Diffusion v1-5 · I2P overall | inappropriate ratio: 0.39 -> 0.09 | Improved |
| apply Geom-Erasing in data removal fine-tuning | SD fine-tuned on ICD-QR · ICD-QR | ICR (%): 74.59 -> 5.38 | Improved |
| apply Geom-Erasing in data removal fine-tuning | SD fine-tuned on ICD-Text · ICD-Text | ICR (%): 71.84 -> 13.48 | Improved |
| erasing implicit concepts also improves FID | SD fine-tuned on ICD-QR · ICD-QR | FID: 65.82 -> 41.41 | Improved |
| use ESD for implicit concept removal | SD fine-tuned on ICD-QR · ICD-QR | FID: 65.82 -> 90.97 | Hurt |
| use FMN or NP or SLD for implicit concept removal | SD · ICD-QR | ICR (%): 74.59 -> 80.42 (FMN); 59.64 (NP) | No clear effect |
| add geometry bins (selected bin number 72 at 32^2) versus smaller or larger bin settings (ablation) | Geom-Erasing · ICD-Watermark | ICR (%): ICR 7.36 default vs 18.26 (8^2) and 16.47 (64^2) | Mixed |
| use detector-confidence re-weighting instead of fixed weight (ablation) | Geom-Erasing · ICD-Watermark | ICR (%): 7.23 -> 9.21 | Hurt |
| add concept name and uniform geometry to negative prompt (ablation) | Geom-Erasing · ICD-Watermark | ICR (%): 7.23 -> 5.02 | Improved |
| add geometric information to negative prompt (ablation) | Geom-Erasing · ICD-Watermark | FID: 6.15 -> 6.99 | Hurt |
| add noise to location tokens (geometry accuracy below 0.4 IoU) (ablation) | Geom-Erasing · ICD-Watermark | ICR | Hurt |
| include the word watermark in the text prompt | Stable Diffusion · watermark presence over 1000 samples | correlation with watermark presence: r=-0.08, p=0.21 | 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 0.89)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 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)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
- Concept Replacer: Replacing Sensitive Concepts in Diffusion Models via Precision Localization (CVPR 2025 · similarity 0.88)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · 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.
