**OpenBias: Open-set Bias Detection in Text-to-Image Generative Models** Text-to-image (T2I) generative models are becoming increasingly popular, but their safety and fairness are crucial concerns. Existing methods focus on detecting closed-set biases, which are well-known and predefined. This paper introduces OpenBias, a novel pipeline that identifies and quantifies biases in T2I models without relying on precompiled sets. OpenBias consists of three stages: (1) a Large Language Model (LLM) proposes biases given captions, (2) the generative model produces images using these captions, and (3) a Vision Question Answering (VQA) model assesses the presence and extent of biases. The study focuses on Stable Diffusion 1.5, 2, and XL, uncovering new biases not previously investigated. Quantitative experiments show that OpenBias agrees with closed-set bias detection methods and human judgment. **Contributions:** - OpenBias is the first large-scale open-set bias detection method. - It discovers novel biases that have not been studied before. - The pipeline is modular and flexible, allowing for component updates. - It treats the generative model as a black box, querying it with specific prompts. **Experiments:** - OpenBias is evaluated on Flickr30k and COCO datasets. - It compares well-known biases with classifier-based methods and human judgment. - A user study shows high alignment between OpenBias and human judgments. **Findings:** - OpenBias identifies both well-known and novel biases. - It reveals context-aware and context-free biases. - Qualitative results highlight biases related to objects, animals, and persons. **Limitations:** - The pipeline relies on the LLM and VQA models, which may have biases. - Context-aware bias analysis is qualitative, with room for systematic study. **Conclusions:** - OpenBias advances bias detection in T2I models, fostering more inclusive AI. - It provides a valuable tool for bias mitigation research.**OpenBias: Open-set Bias Detection in Text-to-Image Generative Models** Text-to-image (T2I) generative models are becoming increasingly popular, but their safety and fairness are crucial concerns. Existing methods focus on detecting closed-set biases, which are well-known and predefined. This paper introduces OpenBias, a novel pipeline that identifies and quantifies biases in T2I models without relying on precompiled sets. OpenBias consists of three stages: (1) a Large Language Model (LLM) proposes biases given captions, (2) the generative model produces images using these captions, and (3) a Vision Question Answering (VQA) model assesses the presence and extent of biases. The study focuses on Stable Diffusion 1.5, 2, and XL, uncovering new biases not previously investigated. Quantitative experiments show that OpenBias agrees with closed-set bias detection methods and human judgment. **Contributions:** - OpenBias is the first large-scale open-set bias detection method. - It discovers novel biases that have not been studied before. - The pipeline is modular and flexible, allowing for component updates. - It treats the generative model as a black box, querying it with specific prompts. **Experiments:** - OpenBias is evaluated on Flickr30k and COCO datasets. - It compares well-known biases with classifier-based methods and human judgment. - A user study shows high alignment between OpenBias and human judgments. **Findings:** - OpenBias identifies both well-known and novel biases. - It reveals context-aware and context-free biases. - Qualitative results highlight biases related to objects, animals, and persons. **Limitations:** - The pipeline relies on the LLM and VQA models, which may have biases. - Context-aware bias analysis is qualitative, with room for systematic study. **Conclusions:** - OpenBias advances bias detection in T2I models, fostering more inclusive AI. - It provides a valuable tool for bias mitigation research.