Ara-Pic: A Framework for Enhancing Arabic Cultural Representation in AI-Generated Images

Recent advancements in text-to-image (TTI) models have demonstrated impressive generative capabilities but often fail to produce culturally relevant images, particularly for underrepresented cultures such as Arabic culture. In this paper, we introduce Ara-Pic, a novel approach that iteratively enhances image generation prompts to improve cultural relevance. Our method leverages the capabilities of vision-language models (VLMs) to culturally enhance prompts based on feedback from the Cultural Relevance Index (CRI), a metric designed to assess the cultural alignment of AI-generated images. By systematically modifying prompts to increase positive cultural amplification while reducing penalties, Ara-Pic optimizes image generation to better reflect authentic Arabic cultural elements. We evaluate our approach by comparing baseline CRI scores of images generated with initial prompts with the scores of those refined through Ara-Pic, demonstrating its effectiveness in increasing CRI scores. Our findings highlight the potential of iterative prompting as well as the importance of quantified cultural evaluation for scalable and automated solutions for improving cultural representation in AI-generated content.