Research map / Image generation / Text-to-image diffusion
Minimalist Concept Erasure in Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes minimalist concept erasure for flow-matching text-to-image models, using only final-output distribution distances, end-to-end backprop through all sampling steps, and learned neuron masks on FLUX.1-schnell.
Problem
Concept erasure in large text-to-image generative models often makes excessive weight changes that hurt overall utility and are not robust to adversarial prompts, especially for rectified flow DiT models that lack cross-attention.
Approach
Formulate erasure using only the distributional distance of final generated outputs (not intermediate steps), derive an MSE-based upper bound for rectified flow models, optimize it end-to-end through all Euler steps with step-wise gradient checkpointing, and learn a hard-discrete neuron mask instead of fine-tuning weights to ablate concept connectivity.
Key results
- On nudity, ours reduces Ring-A-Bell K77 detection to 19% (vs. 82% for original FLUX and 91% for EAP) and MMA-Diffusion to 0.4%.
- On IP characters, ours reaches 10% ACC vs. 100% for original FLUX while keeping SSIM 0.51 (best among methods).
- Ablation on 20 filtered prompts yields 4% ACC with CLIP 0.29 and FID 41.3, versus 28% ACC without filtering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply minimalist concept erasure | FLUX.1-Schnell · Inappropriate Objects | ACC: 81% -> 43% | Helped |
| apply minimalist concept erasure | FLUX.1-Schnell · IP Characters | ACC: 11% -> 10% | Helped |
| apply minimalist concept erasure | FLUX.1-Schnell · Art Styles | ACC: 4% -> 1% | Helped |
| apply minimalist concept erasure | FLUX.1-Schnell · Nudity: Ring-A-Bell K77, K38, and K16 | ASR: 82% -> 19%; 83% -> 16%; 79% -> 12% | Helped |
| apply minimalist concept erasure | FLUX.1-Schnell · Nudity: MMA-Diffusion | ASR: 9% -> 0.4% | Helped |
| apply minimalist concept erasure | FLUX.1-Schnell · Nudity: P4D, I2P, and Normal prompts | ASR: 47% -> 19%; 30% -> 9%; 64% -> 4% | Helped |
| use prompt filtering rather than no filtering (ablation) | Minimalist concept erasure on FLUX.1-Schnell · Nudity | ACC: 28% -> 4% | Helped |
| mask FFN and normalization layers together rather than either alone (ablation) | Minimalist concept erasure on FLUX.1-Schnell · Nudity | ACC: 58% -> 4% (FFN); 28% -> 4% (NORM) | Helped |
| use 20 rather than fewer guidance prompts (ablation) | Minimalist concept erasure on FLUX.1-Schnell · Nudity | ACC: 41% -> 4% (1 prompt); 29% -> 4% (8 prompts); 16% -> 4% (16 prompts) | Helped |
| decrease beta (ablation) | Minimalist concept erasure · Nudity | erasure intensity | Helped |
| increase optimization steps (ablation) | Minimalist concept erasure · Nudity | concept erasure | Helped |
Surprising findings
- Masking FFN and normalization neurons outperforms masking attention modules, and FFN-only masking degrades FID substantially (65.3).
Limitations
- Step-wise gradient checkpointing implementation lacks multi-GPU support, limiting scalability for weight-level masking.
- Post-masking fine-tuning could not be conducted due to implementation constraints.
- Effectiveness depends on prompt filtering and consistent-background prompts; optimal erasure step remains open.
Experimental setup
Datasets: GCC3M LAION 5K GPT-4o-generated concept prompts Ring-A-Bell MMA-Diffusion P4D I2P
Benchmarks: Ring-A-Bell MMA-Diffusion P4D I2P LAION 5K
Metrics: ACC (detection rate) CLIP score FID SSIM attack success rate
Baselines: ESD CA SLD EAP FlowEdit
Builds on: rectified flow Erasing Stable Diffusion (ESD) Concept Ablation Hard-discrete sampling step-wise gradient checkpointing
Closest papers by meaning
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.92)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.91)
- Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation (CVPR 2025 · similarity 0.90)
- All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.90)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.90)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 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.