Structured Semantic Translation and Diffusion-Based Generation of Lingnan Courtyard Garden Spatial Layouts: A Multi-Level Evaluation Framework

Generative AI has advanced architectural and landscape layout generation, but regional gardens still lack formal spatial representations, controllable generation of intrinsic spatial prototypes, and domain-aligned evaluation methods. This study presents a theory-informed proof of concept for structured translation, conditional generation, and evaluation of Lingnan courtyard garden layouts. Based on classical Lingnan garden theory, 63 plans are encoded through five semantic layers: period style, spatial structure, organization mode, courtyard type, and constituent elements. The resulting multimodal dataset pairs normalized color-coded plans with structured prompts. A latent diffusion model fine-tuned using low-rank adaptation supports boundary-constrained generation. A multi-level evaluation framework combines spatial-prototype label ablation, expert relevance assessment, and element-proportion analysis. Empirical results show that structured spatial labels improve generation controllability: the structural-label hit rate reaches 78.8% (vs. 19.4% for element-only controls), and the organizational-label hit rate reaches 52.1% (vs. 13.6%). Structural control is stronger, whereas precise organizational relationships remain difficult to generate. This study establishes a traceable pathway from regional garden knowledge encoding to generative control and evaluation, provides an open and reusable dataset, and offers a methodological reference for the digital preservation of regional architectural and landscape heritage.