GLEaN: A Text-to-image Bias Detection Approach for Public Comprehension
Summary: arXiv:2604.09923v1 Announce Type: new
Abstract: Text-to-image (T2I) models, and their encoded biases, increasingly shape the visual media the public encounters. While researchers have produced a rich body of work on bias measurement, auditing, and mitigation in T2I systems, those methods largely target technical stakeholders, leaving a gap in public legibility. We introduce GLEaN (Generative Likeness Evaluation at N-Scale), a portrait-based explainability pipeline designed to make T2I model biases visually understandable to a broad audience.
Introduction to GLEaN
GLEaN comprises three stages: automated large-scale image generation from identity prompts, facial landmark-based filtering and spatial alignment, and median-pixel composition that distills a model’s central tendency into a single representative portrait. The resulting composites require no statistical background to interpret; a viewer can see, at a glance, who a model ‘imagines’ when prompted with terms such as ‘a doctor’ versus a ‘felon.’
Methodology
The GLEaN pipeline consists of the following stages:
- Image Generation: This stage involves generating a large number of images based on specific identity prompts.
- Facial Landmark Filtering: Here, the generated images are filtered and aligned based on facial landmarks to maintain consistency.
- Median-Pixel Composition: This final step combines the filtered images to create a composite portrait that represents the model’s central tendency.
Findings and Impact
We demonstrated GLEaN on Stable Diffusion XL across 40 social and occupational identity prompts, producing composites that not only reproduce documented biases but also surface new associations between skin tone and predicted emotion. Notably, our findings revealed that GLEaN portraits communicate biases as effectively as conventional data tables, but require significantly less viewing time.
User Study
A between-subjects user study involving 291 participants was conducted to evaluate the effectiveness of GLEaN. The results indicated that participants found the GLEaN portraits to be an accessible and efficient means of understanding biases in T2I models without needing extensive statistical knowledge.
Scalability and Accessibility
One of the significant advantages of GLEaN is that it relies solely on generated outputs. This characteristic allows it to be replicated on any black-box and closed-weight systems without requiring access to model internals. This makes GLEaN a scalable, model-agnostic approach to bias explainability, purpose-built for public comprehension.
Conclusion
GLEaN stands as a pioneering effort to bridge the gap between technical research in bias measurement and public comprehension. By offering a visually intuitive tool for understanding biases in T2I models, GLEaN aims to foster a more informed public discourse around the implications of AI in visual media. The method is publicly available at https://github.com/cultureiolab/GLEaN.