Where Culture Fades: Revealing the Cultural Gap in Text-to-Image Generation
Chuancheng Shi, Shangze Li, Shiming Guo, Simiao Xie, Wenhua Wu, Jingtong Dou, Chao Wu, Canran Xiao, Cong Wang, Zifeng Cheng, Fei Shen, Tat-Seng Chua
2025-12-02
Summary
This paper investigates how well artificial intelligence models that create images from text work when given prompts in different languages, focusing on whether the images reflect the correct cultural context.
What's the problem?
Current text-to-image AI models, while good at making realistic images, often produce results that are either culturally bland or heavily influenced by Western (specifically English) culture, even when asked to create images based on prompts in other languages. The issue isn't that the models *don't know* about other cultures, but rather that they don't properly *use* that knowledge when generating images.
What's the solution?
The researchers figured out which specific parts of the AI model's 'brain' (neurons in certain layers) are responsible for understanding cultural cues. Then, they developed two ways to improve the images: one method boosts the activity of those culture-related neurons when creating an image, and the other method slightly adjusts those specific parts of the model to be more sensitive to cultural information. They tested these methods using a new set of images designed to assess cultural accuracy.
Why it matters?
This research is important because as AI image generators become more common, it's crucial that they can accurately and respectfully represent different cultures. If these models only produce Western-centric images, it reinforces biases and limits their usefulness for a global audience. This work helps make AI more inclusive and culturally aware.
Abstract
Multilingual text-to-image (T2I) models have advanced rapidly in terms of visual realism and semantic alignment, and are now widely utilized. Yet outputs vary across cultural contexts: because language carries cultural connotations, images synthesized from multilingual prompts should preserve cross-lingual cultural consistency. We conduct a comprehensive analysis showing that current T2I models often produce culturally neutral or English-biased results under multilingual prompts. Analyses of two representative models indicate that the issue stems not from missing cultural knowledge but from insufficient activation of culture-related representations. We propose a probing method that localizes culture-sensitive signals to a small set of neurons in a few fixed layers. Guided by this finding, we introduce two complementary alignment strategies: (1) inference-time cultural activation that amplifies the identified neurons without backbone fine-tuned; and (2) layer-targeted cultural enhancement that updates only culturally relevant layers. Experiments on our CultureBench demonstrate consistent improvements over strong baselines in cultural consistency while preserving fidelity and diversity.