Research map / Image generation / Text-to-image diffusion
Sculpting Memory: Multi-Concept Forgetting in Diffusion Models via Dynamic Mask and Concept-Aware Optimization
Gen Li, Yang Xiao, Jie Ji, Kaiyuan Deng, Bo Hui, Linke Guo, Xiaolong Ma
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 dynamic mask instead of fixed mask (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.175 -> 0.084 | Improved |
| add dynamic mask instead of fixed mask (ablation) | Ours · Imagenette (unlearn 3 classes) | Total Acc: 0.194 -> 0.093 | Improved |
| raise cross-attention sparsity from 20% to 50% (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.215 -> 0.084 | Improved |
| raise cross-attention sparsity from 50% to 80% (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.084 -> 0.110 | Hurt |
| restrict sparse updates to K and V matrices (50%) instead of cross-attention layers (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.084 -> 0.209 | Hurt |
| apply sparsity to whole UNet instead of cross-attention (50%) (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc / CLIP: 0.084 / 26.43 -> 0.001 / 19.91 | Mixed |
| mask update every 100 steps instead of 50 (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.120 -> 0.084 | Improved |
| mask update every 100 steps instead of 400 (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.125 -> 0.084 | Improved |
| mask update every 100 steps instead of 200 (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc: 0.095 -> 0.084 | Improved |
| add regularization (distillation) loss with beta 0.25 (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc | Improved |
| increase alignment loss weight alpha beyond 0.25 (ablation) | Ours · Imagenette (unlearn 10 classes) | Total Acc | Improved |
| use Ours instead of MACE on NSFW (I2P) | Ours vs MACE · I2P NudeNet detection (all prompts) | Total NSFW detections: 111 -> 47 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SalUn — SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024) · better on 3 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
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.91)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.90)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.89)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.89)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · 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.
