The paper investigates evaluation methodologies for AI-based virtual reconstruction of historical architecture, focusing on the uncertainty inherent in generative predictions. The research critically compares traditional analog reconstruction workflows with AI-driven reconstruction processes based on GAN networks, proposing a conceptual framework called the “Uncertainty Tree” to describe the decision-making chain involved in reconstruction. The study develops a qualitative evaluation methodology intended to complement quantitative pixel-based assessments by introducing prediction thresholds ranging from impossible to highly probable reconstructions. Through experiments on datasets of Greek Doric temples and Mudejar Romanesque churches, the research argues that specialized datasets and expert qualitative evaluation significantly improve the plausibility and reliability of AI-generated architectural reconstructions. The paper ultimately proposes a new epistemological framework for evaluating AI predictions in heritage reconstruction, emphasizing human supervision, dataset specificity, and iterative qualitative assessment.
Automatic Virtual Reconstruction of Historic Buildings Through Deep Learning. A Critical Analysis of a Paradigm Shift
This paper critically examines the emerging use of artificial intelligence for the automatic virtual reconstruction of historic buildings, comparing it with traditional heritage reconstruction methodologies. The authors trace the historical evolution of reconstruction practices—from manual drawing, archaeological interpretation, CAD modelling, photogrammetry, and BIM—to current AI-based approaches founded on Deep Learning. The study focuses on Generative Adversarial Networks (GANs) trained to infer missing architectural parts from ruined structures, using synthetic datasets of Greek temples represented in multiple ruin states and reconstructed versions. A complementary Natural Language Processing workflow is also tested to improve segmented image generation and automate parts of the training process. Results suggest that AI can identify hidden formal and constructive patterns, generate multiple predictive alternatives, and support specialists in evaluating reconstruction hypotheses. At the same time, the paper warns that neural networks simplify architectural diversity into generalized stylistic rules and still depend on carefully designed datasets. The authors conclude that traditional scholarly expertise and AI prediction systems will likely coexist, marking a paradigm shift in how virtual reconstruction may be practiced in the near future.