Research map / Image generation / Text-to-image diffusion
Exposing Hidden Biases in Text-to-Image Models via Automated Prompt Search
Manos Plitsis, Giorgos Bouritsas, Vassilis Katsouros, Yannis Panagakis
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 |
|---|---|---|---|
| raise classifier weight lambda from 10 to 100 (ablation) | BGPS · Male-biased gender prompts, SD 1.5 | mean male frequency: 0.76 -> 0.91 | Improved |
| raise classifier weight lambda from 10 to 100 (ablation) | BGPS · Female-biased gender prompts, SD 1.5 | mean female frequency: 0.43 -> 0.66 | Improved |
| raise classifier weight lambda from 10 to 100 (ablation) | BGPS · Male-biased prompts, SD 1.5 | prompt perplexity: 53 -> 129 | Hurt |
| add bias-classifier guidance (BGPS lambda=100) to LLM-only decoding | LLM prompts · Female-biased gender prompts, SD 1.5 | mean female frequency: 0.27 -> 0.66 | Improved |
| replace gradient-based PEZ with BGPS | prompt search · Male-biased prompts, SD 1.5 | prompt perplexity: 1387 -> 53 | Improved |
| add BGPS guidance to LLM on underrepresented group | LLM (biased) · Black-biased prompts, SD 1.5 | mean black frequency: 0.05 -> 0.27 | Improved |
| use BGPS prompts for further debias fine-tuning | LoRA-debiased SD 1.5 · Male-biased prompts, lambda=100 | male proportion: 0.71 -> 0.47 | Improved |
| raise lambda from 10 to 100 on age attribute (ablation) | BGPS · Old-biased prompts, SD 1.5 | old frequency: 0.18 -> 0.35 | Improved |
| add BGPS guidance (lambda=10) to LLM-only | LLM prompts · Young/old age prompts | young frequency: 0.81 -> 0.80 | No clear effect |
| increase lambda from 5 to 100 (ablation) | BGPS · Flux male-biased prompts | male frequency: 0.68 -> 0.61 | Hurt |
| apply BGPS to debiased DiffLens model | BGPS (lambda=100) · Male-biased gender prompts, DiffLens | mean male frequency: 0.31 -> 0.67 | Improved |
| halve K from 10 to 5 (ablation) | BGPS · Runtime | minutes per prompt: ~2.4 -> ~1.2 | Improved |
Closest papers by meaning
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.92)
- Implicit Bias Injection Attacks against Text-to-Image Diffusion Models (CVPR 2025 · similarity 0.92)
- Finetuning Text-to-Image Diffusion Models for Fairness (ICLR 2024 · similarity 0.91)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.90)
- OpenBias: Open-set Bias Detection in Text-to-Image Generative Models (CVPR 2024 · similarity 0.89)
- Fair Text-to-Image Diffusion via Fair Mapping (AAAI 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.
