Problem

Large Vision-Language Model (LVLM)-based text-to-image (T2I) systems have become dominant in image generation, but their propensity to amplify social biases remains largely unaddressed. Existing research primarily focused on older architectures, leaving the impact of LVLMs—which actively interpret and reformulate user inputs via hidden transformations—as an underexplored area. The paper hypothesizes that these internal reformulations, particularly through system prompts, are a core source of social bias, leading to the implicit insertion of demographic attributes and contextual assumptions not present in original user prompts.

Method

The authors developed a comprehensive 1,024-prompt benchmark spanning four levels of linguistic complexity, from simple occupations to detailed scene descriptions, to systematically evaluate demographic bias across attributes like age, gender, ethnicity, and appearance. They compared six recent T2I models, including LVLM-based (SANA and Qwen-Image) and non-LVLM models. Bias was quantified using Fair Discrepancy, and text-image alignment with CLIP score. To understand bias mechanisms, they analyzed decoded intermediate representations, token-probability preferences, and embedding-association analyses. Based on these insights, they proposed FairPro, a training-free meta-prompting framework. FairPro leverages the embedded LVLM to self-audit user inputs and dynamically generate fairness-aware system prompts at test time, utilizing chain-of-thought reasoning to mitigate stereotypes.

Results

The study found that LVLM-based T2I models consistently exhibit substantially stronger demographic biases than non-LVLM counterparts, with Qwen-Image and SANA showing the highest bias. Prompts with explicit demographic attributes and LLM-based rewriting techniques were observed to significantly amplify bias. A strong positive correlation (r=0.948) was identified between text-image alignment and social bias, suggesting that improved semantic understanding often comes at the cost of increased stereotypes. Mechanistic analyses revealed that system prompts inject linguistic biases into decoded texts, influence token-level gender preferences (reducing neutrality by 27-36% when removed), and create pronounced gender associations in text embeddings. FairPro effectively reduced bias across all demographic attributes and prompt complexities for SANA and Qwen-Image, while preserving text-image alignment. Ablation studies confirmed the necessity of user context and chain-of-thought reasoning for optimal performance.

Implications

This research offers crucial insights into the pervasive issue of social bias in modern T2I systems, particularly highlighting the previously under-examined role of system prompts in propagating biases within LVLM-based architectures. The introduction of FairPro provides a practical, deployable, and training-free solution for mitigating these biases, enabling the development of more socially responsible generative AI. While FairPro is an input-level intervention, the authors acknowledge limitations such as the reliance on binary perceived-gender annotations and suggest future work could incorporate more inclusive and nuanced attribute annotations to further enhance fairness evaluation.