Controllable AI modeling of northern and southern timbercraft convergence in Qing interior canopies

Abstract

This study examines interior canopies from the Qianlong–Jiaqing era (1736–1820) of the Qing dynasty to model northern–southern timbercraft convergence through controllable AI generation. We constructed northern official and southern garden semantic lexicons, trained dual LoRA modules on Stable Diffusion v1.5, and generated intermediate canopy samples using controlled alpha-based weighting. Structural, ornamental, and semantic indicators were then extracted and compared with historical Yangzhou canopy samples. As southern weighting increased, generated samples showed lower aspect ratios and higher openwork rate, curve complexity, pattern density, decorative layering, and southern tendency scores. PCA and ROI-based comparison indicate that Yangzhou samples are most compatible with the mid-range model space, especially the alpha 0.50–0.75 groups, rather than with either endpoint. The study does not treat AI outputs as historical evidence, but proposes controllable AI as a modeling and compatibility-assessment tool for architectural history and heritage research.