Research map / Image generation / Text-to-image diffusion
Push–Pull Attentional Anchoring for Diffusion Concept Erasure
Nattanat Chatthee, Tagon Sompong, Ekapol Chuangsuwanich, Supasorn Suwajanakorn
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 |
|---|---|---|---|
| use PPAA cosine push-pull erasure instead of ESD | SD1.4 nudity erasure · Nudity erasure, UnlearnDiff attack | ASR: 0.92 -> 0.83 | Improved |
| use PPAA instead of STEREO for nudity erasure | SD1.4 · Nudity erasure, CMMD preservation | CMMD: 139.47 -> 1.43 | Improved |
| use PPAA instead of DUO/STEREO regarding attack robustness | SD1.4 · Nudity erasure, CCE attack | ASR: 0.01 (DUO) / 0.04 (STEREO) vs 0.20 | Hurt |
| use relative pull margin instead of absolute margin (ablation) | PPAA · SD1.4 erasure-preservation | erasure-preservation trade-off | Improved |
| use cosine-based losses instead of L2-based losses (ablation) | PPAA · SD1.4 erasure-preservation | erasure-preservation trade-off | Improved |
| add preservation losses L_preserve and L_denoise (ablation) | PPAA · SD1.4 erasure-preservation | non-target preservation | Improved |
| lower m_push (stronger semantic push) (ablation) | PPAA · SD1.4 target detection | target-concept detection accuracy | Improved |
| apply push loss only to concept tokens (exclude EOS) (ablation) | PPAA · SD1.4 qualitative | localization of semantic change | Improved |
| apply PPAA objective to FLUX | FLUX · FLUX Obama/Van Gogh/Mickey | erasure-preservation trade-off | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- SUMA — SuMa: A Subspace Mapping Approach for Robust and Effective Concept Erasure in Text-to-Image Diffusion Models (ICCV 2025) · better on 1 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.88)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · similarity 0.88)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.87)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.87)
- Easing Concept Bleeding in Diffusion via Entity Localization and Anchoring (ICML 2024 · similarity 0.87)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 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.
