A systematic and procedural framework for digital reconstruction of architectural cultural heritage using generative AI

Cultural heritage remains vulnerable to deterioration and loss, while physical restoration is constrained by principles of minimal intervention, reversibility, and material compatibility. Digital reconstruction, therefore, offers a non-invasive, evidence-based alternative. However, no integrated framework systematically connects target selection, model generation, and interpretation, limiting consistency and transparency. This study proposes the SGAD (Selection–Generation–Assessment–Documentation) framework for procedural coherence, reproducibility, and interpretative accountability. The Selection phase defines reconstruction scope, evidence grading, and practical criteria for disclosure pathways. The Generation phase applies a three-stage workflow—input preparation, model training, and integrated generation—using stable diffusion with element-specific LoRA modules. The Assessment phase evaluates quality through structural–morphological and perceptual–local axes, and the Documentation phase records metadata on hypothesis status, transparency, cultural sensitivity, and interpretative boundaries. Applied to Yeongsanjeon Hall, Magoksa Temple, SGAD demonstrates feasibility and interpretative considerations, supporting a structured approach to architectural digital reconstruction.