Research map / Image generation / Text-to-image diffusion
Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion 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.
Builds a full-pipeline toolkit and benchmark for NSFW concept erasure in text-to-image diffusion models, comparing 13 erasure methods on nudity, violence, and horror with VQA-based detection and a set of erasure and generation metrics.
Problem
Concept erasure methods for suppressing NSFW content in text-to-image diffusion models lack comprehensive, fine-grained evaluation across scenarios, including robustness, data sensitivity, and preservation of unrelated concepts.
Approach
Provides a benchmark with thematically annotated NSFW prompt datasets, a taxonomy of erasure methods (stage, required data, trained component), a VQA-based classifier chosen after comparison with CLIP, NudeNet, and MHSC, and new metrics (EP, GRD) for systematic comparison.
Key results
- Post-hoc SLD-Max achieves the highest overall erasure proportion (91.81%), while SD-NP reaches 71.65%; ESD-x and ESD-u reach only 7.88% and 5.44%.
- On the RAB adversarial prompts, AU achieves 93.96% overall erasure proportion versus 34.25% for SD-NP, while SLD-Max reaches 96.09%.
- VQA achieves 88.62% nudity accuracy and 86.41% recall, outperforming CLIP, MHSC, and NudeNet in recall across themes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase SLD intervention strength from Medium to Strong to Max (ablation) | SLD · I2P, 4chan, Lexica, Template (Overall) | EP: 58.43% -> 80.22% -> 91.81% | Helped |
| Increase SLD intervention strength from Medium to Strong to Max (ablation) | SLD · RAB (Overall) | EP: 32.01% -> 73.59% -> 96.09% | Helped |
| Increase SLD intervention strength from Medium to Strong to Max (ablation) | SLD · COCO-10K | FID: 24.02 -> 27.72 -> 33.43 | Hurt |
| Increase SLD intervention strength from Medium to Strong to Max (ablation) | SLD · COCO-10K | CLIPS: 25.47 -> 24.66 -> 23.75 | Hurt |
| Use concise rather than detailed keyword sets (ablation) | Mode-1 methods · Nudity, violence, and horror | EP | No clear effect |
| Increase image training data scale (ablation) | Mode-2 methods · Nudity, violence, and horror | EP | Mixed |
| Increase image training data scale (ablation) | Concept erasure methods · COCO-10K | Image quality | Hurt |
| Use adversarial training | AU · RAB | EP: 93.96% | Helped |
| Train with image-level suppression using unsafe images and masks | MACE · RAB | EP: 84.99% | Helped |
| Increase SLD intervention strength (ablation) | SLD · Excessive erasure of nudity | GRD: 12.03% -> 5.98% -> 3.81% | Hurt |
| Use unsafe images to identify sensitive parameters and fine-tune | SalUn · Excessive erasure of nudity | GRD: -6.38% | Hurt |
| Increase SLD intervention strength (ablation) | SLD · RAB (Overall) | EP: 32.01% -> 73.59% -> 96.09% | Helped |
Surprising findings
- Blindly increasing the training data scale yields limited benefits for most erasure methods, and larger datasets can degrade image quality.
- Post-hoc methods, though efficient and strong in standard tests, are easily circumvented and thus unsuitable for open-source or high-security settings.
Limitations
- Covers only nudity, violence, and horror themes; excludes hate, extremist, political, and legal topics.
- Relies on a VQA classifier whose judgments can diverge from human labels due to varying tolerance levels and artistic interpretation.
- Evaluation restricted to Stable Diffusion v1.4 based erasure methods.
Experimental setup
Datasets: I2P 4chan Lexica Template COCO-10K
Benchmarks: I2P 4chan Lexica Template COCO-10K RAB
Metrics: Erasure Proportion Genital Ratio Difference FID LPIPS CLIPScore ImageReward accuracy recall F1
Baselines: SD-NP SLD-Med SLD-Str SLD-Max ESD-u ESD-x SPM UCE AU AC SelfD SalUn
Builds on: ESD SPM UCE AU MACE
Closest papers by meaning
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.91)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.90)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.90)
- MACE: Mass Concept Erasure in Diffusion Models (CVPR 2024 · similarity 0.90)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Precise, Fast, and Low-cost Concept Erasure in Value Space: Orthogonal Complement Matters (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.