Causes and Challenges of Bias in Text-to-Image Models

As text-to-image (TTI) generation technology advances rapidly, identifying and addressing the bias inherent in the images models generate is emerging as a critical ethical challenge.

Main Causes of Bias:

• Bias in Training Data: Western-centric values or cultural stereotypes contained in large-scale multimodal datasets such as LAION-5B and MS-COCO are reflected in the models. For example, when asked to generate architecture from a specific region, the results may reflect a Western perspective.
• Bias in the Data Filtering Process: Research findings show that the process of filtering data before training can actually amplify bias against specific groups.
• Reproduction of Social Stereotypes: A tendency appears to repeatedly generate specific races, genders, and cultural characteristics for occupational groups (e.g., CEOs and managers) or cultural symbols (e.g., bridal attire).

Since current TTI models risk amplifying existing social inequalities, there is an urgent need to develop new evaluation methodologies and tools that can accurately measure and mitigate this bias.