Evaluating Occupational Bias in Anime-style Text-to-Image Generation: An MLLM-Based Workflow

Abstract

Gender and age representation biases in photorealistic text-to-image generation have been widely studied, but biases in anime-style image generation remain underexplored. In this study, we propose an MLLM-based workflow for scalable evaluation of gender presentation and apparent age in images of people in occupational settings generated by text-to-image models. The workflow consists of four stages: (1) constructing prompts that combine an occupation and an image style, (2) generating images with the target models, (3) using an MLLM to infer the gender presentation and apparent age of depicted persons and output the results as structured data, and (4) applying statistical tests to identify representational differences across image styles and models. As a case study, we apply the workflow to eleven occupations selected from the OECD ranking of teenagers' occupational aspirations and compare anime-style and photorealistic images generated by five text-to-image models. We find that, compared with photorealistic images, anime-style images often exhibit more extreme female ratios, approaching either 0% or 100%, and skew the apparent-age distribution toward younger age groups. These distributions frequently diverge from both official labor statistics and perceptions elicited from Japanese crowdsourcing participants. We further assess the MLLM labels by comparing them with human annotations and report cases in which generated images closely resemble existing characters or contain elements derived from existing IP.