OpenBias introduces a novel approach for open-set bias detection in text-to-image generative models, without relying on any precompiled set of biases. Instead, it constructs a pipeline utilizing a Large Language Model (LLM) to identify potential biases from captions, which are then used by a generative model to produce images. Subsequently, a Vision Question Answering model quantifies the recognized biases. This methodology not only aligns with human judgment but also concurs with existing closed-set bias detection methods.

• Utilizes LLMs for proposing potential biases from textual data.
• Employs generative models to illustrate biases visually.
• Quantifies biases using Vision Question Answering techniques.
• Extends research beyond predefined bias sets.
• Agrees with human bias detection and established methods.

The significance of OpenBias lies in its ability to identify new, previously unstudied biases, which is crucial for ensuring the fairness and safety of generative models. This work paves the way for more transparent methodologies in AI, encouraging further research into unbiased AI systems and automated bias detection strategies.