Exclusive Flux: A Review of Flux’s Generation of LGBTQ+ Couples

Abstract

The increasing scope and public use of Generative Artificial Intelligence (GenAI) platforms, particularly image generation tools, have prompted questions about the safety and fairness of large Vision Language Models (VLMs), e.g., Flux and DALL-E. The ubiquity and convincing realism of AI-generated imagery injects significant challenges into modern digital literacy efforts because VLMs may unintentionally perpetuate historical stereotypes as a result of biases in training data scraped from the web. Because these VLMs are open-use and their synthetic images are not subject to copyright permissions, these model biases can have far-reaching effects that cement societal biases and reinforce exclusionary practices. Therefore, it is critical to explore and identify bias within these models and to cultivate an understanding of the cultural context in which these biases are echoed as a first step to mitigating these problems. This paper provides an in-depth study of bias against LGBTQ+ individuals in images generated by Flux, the leading image generator model. This work uses a One-Factor-at-a-Time (OFAT) approach to critique the heterosexism present in Flux’s generations, discusses the impact that biased GenAI imagery may have on society, and provides a survey of existing mitigation strategies. The results of these experiments highlight a lack of nuance in Flux’s training, leading to biased synthetic image generation.