Research map / Image generation / Text-to-image diffusion
Localized Concept Erasure in Text-to-Image Diffusion Models via High-Level Representation Misdirection
Uichan Lee, Jeonghyeon Kim, Sangheum Hwang
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 |
|---|---|---|---|
| apply misdirection loss to first-block output (early layer) instead of final block (ablation) | Diff-Q* (early layer random misdirection) · I2P-931 / COCO-1k | LPIPS, CLIP: CLIP 0.273 -> 0.187; I2P 7.09% -> 0.75% | Hurt |
| supervise with final-block high-level representation while updating first block (ablation) | HiRM-R · text encoder erasure | retention vs erasure | Improved |
| use final-layer W_out as target rather than early-layer targets (ablation) | HiRM-R · target-layer ablation | erasure and retention | Improved |
| replace random misdirection with semantic misdirection (ablation) | HiRM-R -> HiRM-S · UnlearnCanvas object erasing | AA: 94.65 -> 96.94 | Improved |
| replace random misdirection with semantic misdirection (ablation) | HiRM-R -> HiRM-S · UnlearnCanvas style erasing | AA: 94.24 -> 95.54 | Improved |
| replace random misdirection with safety vector (ablation) | HiRM-R -> HiRM-S · I2P | ASR (%): 0.96 -> 0.66 | Improved |
| replace random misdirection with safety vector (ablation) | HiRM-R -> HiRM-S · COCO-10k | FID: 8.05 -> 6.75 | Improved |
| add HiRM-R text encoder to ESD | ESD on Flux1.dev · Ring-16 | ASR (%): 41.05 -> 12.63 | Improved |
| add HiRM-R text encoder to EraseAnything | EraseAnything on Flux1.dev · Ring-16 | ASR (%): 29.47 -> 3.16 | Improved |
| add HiRM-R text encoder to CA | CA on Flux1.dev · MMA | ASR (%): 3.20 -> 4.40 | Hurt |
| add HiRM-S text encoder to SPEED U-Net | SPEED · 50-celebrity retention | Acc_r: 88.48 -> 79.30 | Hurt |
| merge six style LoRA modules by plain averaging | HiRM · UnlearnCanvas multi-style | in-domain accuracy: UA 88.33; IRA 65.56; CRA 98.8 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SALUN — SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation (ICLR 2024) · better on 2 of 2 shared results
- EraseAnything — EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025) · better on 1 of 1 shared result
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
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 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)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.91)
- Concept Pinpoint Eraser for Text-to-image Diffusion Models via Residual Attention Gate (ICLR 2025 · similarity 0.91)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.90)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.90)
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.
