Generative Bias Behind the Pixels: Unveiling the Hidden Stereotypes in Text-to-Image Models

Vedat Porikli, Sedat Porikli

Keywords: Text-to-Image, demographic biases and underrepresentation

TL;DR: Our findings reveal significant disparities in T2I generation results across different prompt categories such as actions, attributes, roles, emotions, ideologies, family structures, place descriptions, religion and life events.

Abstract: Text-to-Image (T2I) models have revolutionized image creation by generating highly realistic visuals from textual prompts. While these models are increasingly adopted across various industries, their potential to perpetuate and amplify biases poses significant concerns. This paper explores the biases inherent in T2I models, focusing on gender, racial, age, somatotype and other human-centric factors. Through the generation and analysis of 24,000 images based on 160 prompt topics, we examine the representation of diverse thematic groups across different categories, including actions, attributes, roles, emotions, ideologies, family structures, place descriptions, religion and life events. Our findings reveal significant disparities in generated images, often reinforcing harmful societal stereotypes. We discuss the implications of these biases and advocate for more inclusive datasets to mitigate these issues.

Submission Number: 5