Research map / Image generation / Text-to-image diffusion
SDErasure: Concept-Specific Trajectory Shifting for Concept Erasure via Adaptive Diffusion Classifier
Fengyuan Miao, Shancheng Fang, Lingyun Yu, Yadong Qu, Yuhao Sun, Xiaorui Wang, 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 |
|---|---|---|---|
| add early-preserve loss L_p (ablation) | SDErasure · Celebrity batch erasure | MSCOCO-30k FID: 8.05 -> 7.39 | Improved |
| add concept-retain loss L_r (ablation) | SDErasure · Celebrity batch erasure | MSCOCO-30k FID: 8.05 -> 7.06 | Improved |
| add both L_p and L_r quality regulation (ablation) | SDErasure · Celebrity batch erasure | MSCOCO-30k FID: 8.05 -> 6.65 | Improved |
| add both L_p and L_r quality regulation (ablation) | SDErasure · Celebrity batch erasure | CLIP (lower = more erased): 10.27 -> 10.91 | Hurt |
| use heuristic similar anchor instead of random generic anchor (ablation) | SDErasure · CIFAR-10 object erasure | harmonic mean Ho: 90.21 -> 95.33 | Improved |
| use heuristic similar anchor instead of empty anchor (ablation) | SDErasure · CIFAR-10 object erasure | harmonic mean Ho: 88.17 -> 95.33 | Improved |
| use heuristic anchor instead of random anchor (ablation) | SDErasure · CIFAR-10 object erasure | unaltered accuracy Acc_s: 92.96 -> 98.00 | Improved |
| sparse SSScore sampling every 5 steps instead of every step (ablation) | SDErasure · Artist style erasure | pre-computation time: 163.5s -> 34s | Improved |
| sparse SSScore sampling every 5 steps instead of every step (ablation) | SDErasure · Artist style erasure | MSCOCO FID: 8.06 -> 8.52 | Hurt |
| sparse SSScore sampling every 3 steps instead of every step (ablation) | SDErasure · Artist style erasure | target CLIP: 20.98 -> 18.60 | Improved |
| select timesteps by SSScore threshold (lambda 0.8) instead of all timesteps (lambda 0) (ablation) | SDErasure · Concept erasure averaged | FID and CLIP | Improved |
| select timesteps by SSScore instead of 5 random timesteps (ablation) | SDErasure · Concept erasure averaged | FID | Improved |
Closest papers by meaning
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.94)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.92)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.92)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (CVPR 2025 · similarity 0.92)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.92)
- When Are Concepts Erased From Diffusion Models? (NeurIPS 2025 · similarity 0.92)
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.
