Research map / Image generation / Text-to-image diffusion
Beautiful Images, Toxic Words: Understanding and Addressing Offensive Text in Generated Images
Aditya Kumar, Tom Blanchard, 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 |
|---|---|---|---|
| restrict fine-tuning to text-generation layers instead of all attention layers (ablation) | NSFW-Intervention (SD3) · ToxicBench NSFW words (test) | delta NGramLD (NSFW): +0.49 -> +3.63 | Improved |
| restrict fine-tuning to text-generation layers instead of all attention layers (ablation) | NSFW-Intervention (SDXL) · ToxicBench NSFW words (test) | delta NGramLD (NSFW): 1.97 -> 5.69 | Improved |
| restrict fine-tuning to text-generation layers instead of all attention layers (ablation) | NSFW-Intervention (DeepFloyd IF) · ToxicBench NSFW words (test) | delta NGramLD (NSFW): 0.19 -> 4.72 | Improved |
| fine-tune the diffusion backbone rather than the CLIP text encoder (ablation) | NSFW-Intervention (SD3) · ToxicBench NSFW words (test) | NGramLD NSFW minus benign delta: 1.40 -> 2.10 | Improved |
| apply NSFW-Intervention to SD3 | SD3 · ToxicBench NSFW words (test) | NGramLD: 1.84 -> 5.47 | Improved |
| apply NSFW-Intervention to SD3 benign words | SD3 · ToxicBench benign words | NGramLD: 2.27 -> 4.34 | Hurt |
| apply NSFW-Intervention to SDXL | SDXL · ToxicBench NSFW words (test) | NGramLD: 2.11 -> 7.80 | Improved |
| apply NSFW-Intervention to DeepFloyd IF | DeepFloyd IF · ToxicBench NSFW words (test) | NGramLD: 1.85 -> 6.57 | Improved |
| apply NSFW-Intervention to NSFW prompts in user study | SD3 · User study NSFW prompts | recognition rate (%): 78.67 -> 26.56 | Improved |
| apply NSFW-Intervention to misspelled NSFW prompts | SD3 · User study misspelled NSFW | recognition rate (%): 76.41 -> 11.45 | Improved |
| apply NSFW-Intervention to benign prompts | SD3 · User study benign prompts | recognition rate (%): 83.43 -> 55.40 | Hurt |
| evaluate on 2.4x longer prompt templates | NSFW-Intervention (SD3) · ToxicBench longer prompts | NGramLD NSFW: 2.43 -> 6.64 | Improved |
Closest papers by meaning
- Comprehensive Assessment and Analysis for NSFW Content Erasure in Text-to-Image Diffusion models (NeurIPS 2025 · similarity 0.89)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.89)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.88)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.88)
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models (CVPR 2025 · similarity 0.87)
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.
