Research map / Image generation / Text-to-image diffusion
Locality-Aware Continual Unlearning for Diffusion Models
Naveen George, Naoki Murata, Yuhta Takida, Konda Reddy Mopuri, Yuki Mitsufuji
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 local replay (L_retain) to unlearning loss (ablation) | LACU · Loss ablation (10 steps) | RR_acc: 0.32 -> 0.61 | Improved |
| add local replay (L_retain) to unlearning loss (ablation) | LACU · Loss ablation (10 steps) | GR_acc: 0.67 -> 0.85 | Improved |
| add l2 parameter regularization to unlearning loss (ablation) | LACU · Loss ablation (10 steps) | RR_acc: 0.32 -> 0.39 | Improved |
| add l2 regularization on top of unlearning plus replay (ablation) | LACU · Loss ablation (10 steps) | RR_acc: 0.61 -> 0.74 | Improved |
| add l2 regularization on top of unlearning plus replay (ablation) | LACU · Loss ablation (10 steps) | GR_acc: 0.85 -> 0.87 | Improved |
| add replay and regularization to unlearning only (ablation) | LACU · Loss ablation (10 steps) | U_acc: 0.94 -> 0.92 | Hurt |
| rank targets/neighbors by score-prediction distance instead of CLIP (ablation) | LACU · CUL selection variants (10 steps) | RR_acc: 0.61 -> 0.74 | Improved |
| rank targets/neighbors by score-prediction distance instead of CLIP (ablation) | LACU · CUL selection variants (10 steps) | GR_acc: 0.78 -> 0.87 | Improved |
| rank by score-prediction distance instead of direct LLM replacement (ablation) | LACU · CUL selection variants (10 steps) | RR_acc: 0.60 -> 0.74 | Improved |
| use score-based selection instead of naive fixed anchor (ablation) | LACU · CUL selection variants (10 steps) | RR_acc: 0.41 -> 0.74 | Improved |
| use LACU instead of AGE adaptive targets only | AGE · CUL 10 steps | RR_acc: 0.15 -> 0.72 | Improved |
| use LACU instead of strongest-erasure ESD-u | ESD-u · CUL 10 steps | GR_acc: 0.15 -> 0.87 | Improved |
Closest papers by meaning
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.90)
- The Illusion of Unlearning: The Unstable Nature of Machine Unlearning in Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.90)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.90)
- Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models (NeurIPS 2026 · similarity 0.89)
- CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.89)
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.
