An evidence-informed fine-tuning framework for traceable generative redesign of architectural heritage ornaments

Generative heritage redesign requires more than visually plausible synthesis: documented evidence must become computable conditions that guide, constrain and evaluate outputs. This study proposes an evidence-informed fine-tuning framework for heritage ornament redesign. The framework incorporates an evidence-to-conditioning translation layer that converts various source-audited evidence into ontology-linked minimal semantic units (MSUs), structured prompt fields, exclusion fields, and relation-check configurations. A representational five-civilisation corpus is created to support controlled vocabulary, MSU, and relation construction, while validation uses two structurally contrasting ornament families: Chinese eave tiles and Mesopotamian frieze ornaments. Across 700 outputs, the proposed pipeline achieved an expert-rated mean of 4.66, 0.982 SSIM and 0.041 LPIPS with 1.6% (14.0 M) trainable parameters of the base model and 2.5 h total training time, outperforming selected baselines. It is shown to support traceable plausible redesign within explicit evidence, relation, and expert-evaluation boundaries.