GradBias: Unveiling Word Influence on Bias in Text-to-Image Generative Models

Area of Science:

• Artificial Intelligence
• Computer Vision
• Natural Language Processing

Background:

• Text-to-image (T2I) models generate high-quality images but raise concerns about fairness and safety due to potential biases.
• Existing bias detection methods focus on closed sets of predefined biases, limiting their scope.

Purpose of the Study:

• To propose a general framework for identifying, quantifying, and explaining biases in T2I models within an open-set setting.
• To develop methods that do not require a predefined list of biases.

Main Methods:

• A pipeline leveraging a Large Language Model (LLM) to propose potential biases from captions.
• Image generation using the target T2I model with generated captions.
• Bias evaluation using Vision Question Answering (VQA).
• Two framework variations: OpenBias for detection/quantification and GradBias for prompt word contribution analysis.

Main Results:

• OpenBias effectively detects known and novel biases across people, objects, and animals, aligning with existing methods and human judgment.
• GradBias reveals that neutral words can significantly impact biases.
• Both OpenBias and GradBias outperform several baseline methods, including state-of-the-art foundation models.

Conclusions:

• The proposed open-set framework offers a more comprehensive approach to bias detection in T2I models.
• Understanding prompt word influence is crucial for mitigating unintended biases.
• This work contributes to developing fairer and safer generative AI technologies.