TLDR: This research paper investigates gender bias in six state-of-the-art text-to-image models (HunyuanImage 2.1, HiDream-I1-dev, Qwen-Image, FLUX.1-dev, Stable-Diffusion 3.5 Large, Stable-Diffusion-XL) by generating images for five hospital professions (cardiologist, hospital director, nurse, paramedic, surgeon) using various portrait qualifiers. The study found systematic occupational stereotypes, with nurses consistently depicted as women and surgeons predominantly as men across all models. While FLUX.1-dev showed a female skew, other models reinforced male dominance. Crucially, prompt qualifiers like “corporate” reinforced male depictions, while “beautiful” favored female ones, demonstrating that linguistic framing significantly influences demographic outcomes. The findings highlight the urgent need for bias-aware design, balanced defaults, and user guidance in generative AI to prevent the amplification of occupational stereotypes.
Text-to-image (TTI) models, which are increasingly used in various fields from education to creative design, often reflect and amplify societal biases. A recent study delves into this issue, specifically examining gender representation in images generated for hospital professions. The research, titled “Beyond the Prompt: Gender Bias in Text-to-Image Models, with a Case Study on Hospital Professions”, was conducted by Franck Vandewiele, Rémi Synave, Samuel Delepoulle, and Rémi Cozot.
The study investigated six prominent open-weight text-to-image models: HunyuanImage 2.1, HiDream-I1-dev, Qwen-Image, FLUX.1-dev, Stable-Diffusion 3.5 Large, and Stable-Diffusion-XL. The researchers generated 100 images for each combination of five hospital-related professions—cardiologist, hospital director, nurse, paramedic, and surgeon—and five different portrait qualifiers, such as “corporate” or “beautiful”. This meticulous approach allowed for a detailed analysis of how both the profession and the descriptive terms in a prompt influence the gender of the generated images.
Systematic Stereotypes Revealed
The findings revealed a consistent pattern of occupational stereotypes across all models. For instance, nurses were almost exclusively depicted as women, while surgeons were predominantly shown as men. This aligns with traditional societal gender roles, but it also highlights how AI models can reinforce these existing biases, sometimes even exaggerating them compared to real-world demographics.
However, the study also found interesting differences between the models. Qwen-Image and Stable-Diffusion-XL, for example, showed a rigid male dominance in many roles. HiDream-I1-dev presented mixed results, while FLUX.1-dev stood out by skewing female in most roles, a contrast to the other models. HunyuanImage 2.1 and Stable-Diffusion 3.5 Large also reproduced gender stereotypes but showed varying degrees of sensitivity to how the prompt was phrased.
The Impact of Prompt Qualifiers
A key insight from the research was the significant role of “portrait qualifiers” in shaping gender balance. Terms like “corporate” tended to reinforce male depictions, suggesting an implicit association between professionalism and masculinity in the models’ training data. Conversely, qualifiers such as “beautiful” often favored female representations, linking aesthetics more strongly with femininity. The sensitivity to these qualifiers varied greatly among the models; Qwen-Image remained largely unaffected, while FLUX.1-dev, Stable-Diffusion-XL, and Stable-Diffusion 3.5 Large showed a strong dependence on prompt wording.
For example, in FLUX.1-dev, the percentage of male surgeons generated could swing from as low as 5% with the “beautiful” qualifier to 67% with “corporate”. Similarly, for paramedics in Stable-Diffusion 3.5 Large, male representations ranged from 21% (“beautiful”) to 94% (“corporate”). This demonstrates that the specific words used in a prompt can dramatically alter the demographic outcome of the generated image.
Also Read:
- Evaluating Text-to-Image Model Robustness with Logically Equivalent Prompts
- Fairness Under Scrutiny: How Minor Prompt Changes Uncover Bias in RAG Systems
Ethical Implications and Future Directions
The researchers emphasize that these biases are not merely reflections of existing social inequalities but can actively shape perceptions of professional identity. If biased AI-generated images are used in educational materials, recruitment tools, or public communications, they risk legitimizing and perpetuating occupational imbalances. This raises significant ethical concerns regarding fairness and representation in AI-generated content.
To address these issues, the study calls for several mitigation strategies. These include designing models with balanced generation defaults, offering prompt suggestion systems that encourage neutral or diversity-promoting language, and empowering users to explicitly specify demographic attributes to counteract default biases. The paper concludes by stressing the need for bias-aware design and user guidance to prevent the reinforcement of occupational stereotypes in generative AI, suggesting future research should also explore intersectional biases like ethnicity, age, and disability.
