Abstract
<title>Abstract</title> <p>Generative systems can create plausible heritage scenes, but their evidential status remains weak when cultural prompts, structural references, external three-dimensional generation and downstream visualization are mixed in one workflow. This study frames Miao Dong timber scene generation as a Semantic Structure Asset Transfer problem. We combine a scene morphology semantic graph, LoRA05 visual prior, photogrammetry-derived mixed Guizhou Qingyan structural references, Marble and Broker asset preparation, blinded fixed-view evaluation and provenance logging. Evidence is separated across semantic priors, two-dimensional structural conditioning, paired scene block transfer, Qingyan boundary audit and failure state provenance tests. LoRA05 with MASG gave the strongest semantic prior; lineart gave the best structure style balance; and structure-guided C4 outperformed text-only, generic reference and cultural reference strategies. The workflow does not claim survey-grade reconstruction. It makes generated heritage scenes reviewable as bounded semantic, structural, asset and provenance evidence.</p>