Cultural Bias in Text-to-Image Models: A Systematic Review of Bias Identification, Evaluation, and Mitigation Strategies

Despite their continuous advancements, text-to-image (TTI) models often reflect and reinforce cultural biases, perpetuating stereotypes often inherent in their training data. This systematic review critically examines cultural bias in text-to-image (TTI) models, addressing gaps in existing research by analyzing its manifestations, evaluation methods, and mitigation strategies—both directly and through the lens of intersectionality with other bias dimensions. A comprehensive literature review was conducted across multiple major databases, following a rigorously structured search strategy, resulting in the selection of 58 studies spanning bias analysis, evaluation frameworks, and mitigation techniques. Thematic findings highlight that gender bias was the most extensively studied, appearing in 53 studies (91%), followed by racial/ethnic bias (42 studies) and other social biases (41 studies). Furthermore, the review explores how these biases intersect and compound in AI-generated imagery, shaping and reinforcing cultural bias. Our findings reveal the following key aspects: 1) the lack of standardization and scalability in bias evaluation, 2) the lack of a fully effective mitigation strategy, 3) contributed TTI benchmarks favoring Western-centric perspectives. We finally propose future directions to improve fairness and representation in TTI models.