Generative AI for Historical Reconstruction of Architectural Heritage

Generative AI, particularly generative adversarial networks (GANs) and diffusion models, is increasingly used to propose missing parts of historic buildings and districts. This systematic literature review, following PRISMA 2020, examines how these models are applied to the historical reconstruction of architectural heritage, which data and evaluation methods they use, and where evidence is limited. From 4,160 Scopus records (2021–2026), a keyword filter retained 324 records, and title and abstract screening yielded 84 included studies. Most were published in 2025–2026 and concern buildings, facades, or architectural elements. Diffusion models and GANs each appear in 34 of the 84 studies. Named fidelity metrics (FID, SSIM, PSNR, LPIPS) appear in 12 abstracts, and expert or user assessment in 20 others. Few studies test whether a reconstruction is historically accurate, and none concerns Indonesia or Southeast Asia. Digital reconstruction can support heritage protection under SDG 11.4 with less physical intervention, but computing costs are not reported. Findings are preliminary because they rest on titles and abstracts.