TurathiAI: Diffusion-Based Text-to-Image model for regional residential cultural heritage architectural building design

This is the authors' abstract. We don't add key points for this paper.

The integration of innovative technologies like artificial intelligence (AI) in cultural preservation is becoming a notable challenge in the modern era. The rising popularity of modern design trends threatens to erode and replace regional practices and heritage, especially in the eastern context. This study implements a novel multimodal AI approach for local heritage-inspired architectural design and classification. A comprehensive training dataset is curated by capturing photos of local residential buildings in the United Arab Emirates (UAE), selected as a case study of the prominent influence of globalization and rapid urbanization in the east. The dataset curation methodology consists of photo acquisition, image annotation, and building segmentation. The dataset consists of 1,902 samples, distributed across 634 sets of three item types: captured building images, textual annotations, and building masks. A novel automated segmentation approach, based on the Segment Anything Model, was implemented to generate the masks in the dataset. Diffusion-based fine-tuning and low-rank adaptation techniques are explored for fusing traditional and modern architectural design principles. The generative AI models are evaluated using various image quality assessment (IQA) metrics and the results are validated for statistical significance. The trained models and annotated dataset support aligning modern architectural design with regional identity. Future research should explore multimodal evaluation benchmarks by domain experts to evaluate the cultural authenticity of generated designs, in addition to automated building dataset annotation using large language models.