Generative AI for architectural heritage conservation: expert-guided feature encoding, stylistic regeneration and typical application

Digital preservation of architectural heritage is shifting from geometric recording toward semantic regeneration. Traditional photogrammetry and laser scanning achieve high-precision documentation but fall short in interpreting and reproducing the complex grammars of traditional architecture—a limitation especially acute in Chinese timber buildings, with their intricate components (Dougong) and distinct regional dialects. To address this, we establish a systematic workflow integrating AI-Generated Content (AIGC) technologies for expert-guided feature encoding and stylistic regeneration. Using the Qiqushan Great Temple in Sichuan as a case study, we build a closed-loop pipeline combining ChatGPT-5.1 for semantic filtering, Midjourney v6.1 for style exploration, and Stable Diffusion 3.5 fine-tuned via Low-Rank Adaptation (LoRA) for morphological synthesis. From a verified dataset of 148 annotated images, four element categories—eave decorations, arch structures, interior beams, and pattern motifs—are regenerated and evaluated across nine control weights (s = 0.1–1.5). Analysis reveals an optimal range of s = 0.6–0.8, achieving a peak Edge Alignment Score of 0.91 for silhouette-dominant elements. These findings validate AIGC for heritage conservation and motivate Generative Digital Twins, extending static archiving toward generative cultural inheritance, with a three-tier framework defining its applicability and limits.