Research map / Image generation / Text-to-image diffusion
Concept Removal for Frontier Image Generative Models
Aditya Kumar, Pierre Joly, Adam Dziedzic, Franziska Boenisch
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 transcoder concept removal instead of UCE | Flux · UnlearnCanvas style removal | UA (%): 67.43 -> 88.60 | Improved |
| use transcoder concept removal instead of UCE | SD3.5 · UnlearnCanvas object removal | UA (%): 84.45 -> 68.00 | Hurt |
| use transcoder concept removal instead of UCE | SD3.5 · UnlearnCanvas object removal | IRA (%): 82.23 -> 93.00 | Improved |
| use transcoder concept removal instead of LOCOEDIT | Infinity-2B · UnlearnCanvas object removal | UA (%): 85.72 -> 91.20 | Improved |
| use transcoder concept removal instead of UCE/LOCOEDIT | Infinity-2B · FID on non-target data | FID: UCE 77.56 vs ours 90.92 | Mixed |
| use transcoder concept removal instead of baselines under Ring-A-Bell | SD3.5 · Ring-A-Bell | attack success rate (%): 58.08 -> 25.21 | Improved |
| use transcoder concept removal under Ring-A-Bell | Infinity-8B · Ring-A-Bell | attack success rate (%): 56.34 -> 19.78 | Improved |
| use transcoder concept removal instead of UCE under P4D | SD3.5 · P4D | attack success rate (%): 41.15 -> 35.61 | Improved |
| use transcoder concept removal instead of UCE on I2P nudity | Flux · I2P | total inappropriate images: 173 -> 153 | Improved |
| remove 10 concepts sequentially with transcoder instead of iterative weight edits | Transcoder removal · multi-concept style removal | average score | Improved |
| replace ordered bijection pairing with random pairing (ablation) | Transcoder removal · UnlearnCanvas | UA/IRA/CRA: 69.60 -> 69.57 | No clear effect |
| evaluate on unseen styles with transcoder trained on 5 of 10 styles (ablation) | Transcoder removal · UnlearnCanvas style removal (SD3.5) | UA (%): 71.18 -> 52.63 | Hurt |
Closest papers by meaning
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.90)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.89)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.89)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.89)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.88)
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.
