Knowledge-Augmented Stable Diffusion Model for Semantic Representation of Architectural Heritage: A Case Study of Huizhou Traditional Dwellings

Efficiently generating traditional architectural models with regional characteristics is a growing challenge in digital heritage preservation. Existing approaches often lack semantic depth and struggle to integrate construction logic, stylistic rules, and multi-source data. This study addresses these issues through a Knowledge-Augmented Stable Diffusion model, focusing on Huizhou traditional dwellings as a representative case. We propose a hybrid framework that combines Graph-Based Reasoning (GBR) and Stable Diffusion to enable semantic representation and controlled façade generation. First, the compositional features of Huizhou traditional dwellings (HzTD), hierarchical layout, materials, and ornaments, are analyzed and encoded into a shape grammar. Then, using the Property Graph (PG) data model standard and Neo4j, we construct a formal knowledge base that supports reasoning, retrieval, and prompt generation. A Multimodal Diffusion Transformer (MMDiT) is introduced to incorporate KAR-guided ControlNet instances, improving generation accuracy and semantic alignment. Experiments on the façades of Xidi Ancient Village show that the proposed model outperforms baseline generative methods in style consistency, efficiency, and structural fidelity. This research demonstrates the potential of integrating domain knowledge into generative AI for architectural heritage, offering new possibilities for cultural preservation, design assistance, and virtual reconstruction. Future work will explore multimodal 2D-to-3D generation and dynamic scene synthesis.