BAFIS: Dataset + Framework to assess occupational Bias and Human Preference in modern Text-to-image Models

2026-06-18

AI Summary

This study addresses the significant gender and racial biases exhibited by contemporary text-to-image models in generating occupation-related imagery, which conventional evaluation metrics often fail to capture due to their inability to reflect human subjective judgments of fairness. To bridge this gap, the authors propose BAFIS—a fairness evaluation framework that integrates human preference feedback with multilingual prompts—and construct a dataset of 21,140 images aligned with official employment statistics. Using this framework, they systematically assess occupational bias, image quality, and prompt alignment across leading models including Midjourney v6.1, Stable Diffusion 3 Medium, and DALL·E 3. The work pioneers the incorporation of human preference annotations into bias evaluation, revealing systematic disparities in model outputs and demonstrating only partial correlation between human feedback and traditional automated metrics, thereby underscoring the critical role of human judgment in developing equitable text-to-image generation systems.

Abstract

Generative artificial intelligence has the potential to improve productivity and transform the production of creative content. However, existing research indicates that image generation models are significantly influenced by biases. This work investigates the inherent biases and language-induced biases present in text-to-image models within the context of occupation-related image generation, complementing established metrics with human preference feedback. We present a comprehensive evaluation of five current text-to-image models: Midjourney v6.1, Stable Diffusion 3 Medium, DALL-E 3, Playground v2.5, and FLUX.1-dev , focusing on gender and ethnicity bias, image quality, and prompt alignment. To facilitate this evaluation, we developed the "Battle-Arena for Fair Image Synthesis" (BAFIS), a platform designed to collect human feedback on bias in generated images. Furthermore, we created a dataset comprising 21,140 synthetic images generated using multilingual prompts, which serves as a basis for our analysis. We further place our results within a broader social context by comparing them to official statistics from the German Federal Employment Agency. Our findings reveal systematic biases in text-to-image models, with established evaluation metrics in partial correlation with subjective user ratings. Thus, our research emphasizes the need for including human preferences to develop fairer and more inclusive text-to-image models.

Problem

• occupational bias
• text-to-image models
• human preference
• gender bias
• ethnicity bias

Innovation

• text-to-image generation
• occupational bias
• human preference feedback
• fairness evaluation
• multilingual prompts