Research map / Image generation / Text-to-image diffusion
GradBias: Unveiling Word Influence on Bias in Text-to-Image Generative Models
Moreno D’Incà, Elia Peruzzo, Massimiliano Mancini, Xingqian Xu, Humphrey Shi, Nicu Sebe
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 |
|---|---|---|---|
| use GradBias instead of random word selection | word-influence ranking · Word-influence dataset (SD-1.5) | top-1 accuracy: +12.29% | Improved |
| use GradBias instead of random word selection | word-influence ranking · Word-influence dataset (SD-2) | top-1 accuracy: +10.70% | Improved |
| use GradBias instead of random word selection | word-influence ranking · Word-influence dataset (SD-XL) | top-1 accuracy: +6.71% | Improved |
| use GradBias instead of the second-best baseline | word-influence ranking · Word-influence dataset (SD-1.5, SD-2, SD-XL) | top-1 accuracy: +9.58%, +3.28%, +2.82% | Improved |
| add image context via Llava-1.5-13B as answer-ranking baseline | LLM answer-ranking baseline · Word-influence dataset (SD-2) | top-1 accuracy: +7.42% | Improved |
| use Llama3-70B yes/no ranking instead of random | random baseline · Word-influence dataset (SD-1.5, SD-XL) | top-1 accuracy: +2.71% and +3.64% | Improved |
| use dependency-tree ranking instead of random | random baseline · Word-influence dataset (SD-1.5) | top-1 accuracy: 36.36 -> 36.86 | Improved |
| compute the loss at more than only the final denoising step (every 1, 2, 5 or 10 steps vs only the 50th) (ablation) | GradBias · Word-influence dataset (SD-2) | top-1 and top-2 accuracy: -4.55% top-1 and -8.84% top-2 when only the 50-th step is used | Improved |
| swap CLIP-L VQA for Llava-1.5-13B (ablation) | GradBias · Word-influence dataset (SD-2) | top-1 accuracy: 47.98 -> 48.23 | No clear effect |
| swap CLIP-L VQA for Llava-1.5-13B (ablation) | GradBias · Word-influence dataset (SD-XL) | top-1 accuracy: 44.25 -> 43.22 | No clear effect |
| increase number of generated images from N=1 to N=10 (ablation) | GradBias · Word-influence dataset (SD-2) | top-1 accuracy: 47.47 -> 47.98 | No clear effect |
| increase number of generated images from N=1 to N=10 (ablation) | GradBias · Word-influence dataset (SD-XL) | top-1 accuracy: 42.19 -> 44.25 | Improved |
Closest papers by meaning
- OpenBias: Open-set Bias Detection in Text-to-Image Generative Models (CVPR 2024 · similarity 0.95)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.89)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.88)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.88)
- Joint Vision-Language Social Bias Removal for CLIP (CVPR 2025 · similarity 0.87)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · 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.
