Research map / Image generation / Text-to-image diffusion
CGCE: Classifier-Guided Concept Erasure in Generative Models
Viet Nguyen, Vishal Patel
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 token importance weighting s to the gradient refinement (ablation) | CGCE · Nudity removal, I2P (SD-v1.4) | ASR (%): 6.98 -> 4.62 | Improved |
| add token importance weighting s to the gradient refinement (ablation) | CGCE · Nudity removal, UDA (SD-v1.4) | ASR (%): 40.14 -> 27.46 | Improved |
| add token importance weighting s to the gradient refinement (ablation) | CGCE · Nudity removal, COCO | FID: 16.67 -> 16.10 | Improved |
| raise detection threshold tau from 0.50 to 0.75 (ablation) | CGCE · Nudity removal, UDA (SD-v1.4) | ASR (%): 27.46 -> 28.87 | No clear effect |
| lower detection threshold tau from 0.50 to 0.25 (ablation) | CGCE · Nudity removal, SixCD (SD-v1.4) | ASR (%): 2.47 -> 2.28 | No clear effect |
| increase step size eta from 1.0 to 1.5 (ablation) | CGCE · Nudity removal (SD-v1.4) | ASR and generative quality | Mixed |
| use CGCE instead of SAFREE (token-level projection) | SAFREE · Nudity removal, RAB (SD-v1.4) | ASR (%): 74.74 -> 3.87 | Improved |
| use CGCE instead of SAFREE on FLUX with implicit story prompts | SAFREE · Long implicit story prompts, FLUX | ASR (%): 67.00 -> 8.00 | Improved |
| apply CGCE to the base model on FLUX story prompts | FLUX · Long implicit story prompts, FLUX | ASR (%): 78.00 -> 8.00 | Improved |
| apply three concept classifiers jointly | FLUX · Multi-concept (nudity, Van Gogh, church) | ASR (%): 77.00 -> 15.00 | Improved |
| apply CGCE to multi-artist erasure (100 styles) | SD-v1.4 · 100 artistic styles | CLIP_d: 5.99 -> 6.75 | Improved |
| use CGCE instead of fine-tuning (ESD) for training cost | ESD · Efficiency, SD-v1.4 | training time (s): ~4500 -> ~440 | Improved |
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.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 7 of 7 shared results
- EraseFlow — EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025) · better on 6 of 6 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
- Minimalist Concept Erasure in Generative Models (ICML 2025 · similarity 0.92)
- Circumventing Concept Erasure Methods For Text-To-Image Generative Models (ICLR 2024 · similarity 0.91)
- EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment (NeurIPS 2025 · similarity 0.90)
- ConceptPrune: Concept Editing in Diffusion Models via Skilled Neuron Pruning (ICLR 2025 · similarity 0.89)
- Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Models (NeurIPS 2024 · similarity 0.88)
- STEREO: A Two-Stage Framework for Adversarially Robust Concept Erasing from Text-to-Image Diffusion 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.
