Pasar al contenido principal

Resultados relacionados

Resultados totales (Incluyendo duplicados): 1
Encontrada(s) 1 página(s)

RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Can generative AI be a tool for geometric documentation of lost heritage?, ¿Puede la IA generativa ser una herramienta de documentación geométrica del patrimonio desaparecido?

  • Fiz, Ignacio
  • Martín, Pere Manel
  • Hofman, Vanina Yael
  • Puente, Ginés
  • Moruno, Carlos
  • Mallafrè, Celia
[EN] This paper examines the application of Generative Artificial Intelligence (GenAI) to the recovery of lost architectural heritage, assessing both its visual potential and its current technical limitations. The emergence of GenAI in 2023 has had a significant impact on scientific and academic fields, generating debates reminiscent of Umberto Eco's distinction between the "apocalyptic" and the "integrated." Beyond this philosophical discussion, this study evaluates whether GenAI may represent a new paradigm for the geometric documentation of lost heritage, particularly in cases where only a small number of photographs or illustrations survive as a result of urban transformation or armed conflict. The aim is to explore whether these tools can overcome the limitations of traditional "reverse engineering" methods when available information is extremely scarce. To date, the creation of "digital twins" has relied mainly on photogrammetry, including Structure from Motion (SfM) and Multi-View Stereo (MVS) algorithms, as well as three-dimensional (3D) scanning. These techniques provide high levels of accuracy but require extensive data and specialist training. Although Neural Radiance Fields (NeRFs) have enabled photorealistic representations by modelling light and radiance, they still depend on multiple viewpoints. GenAI introduces new possibilities through Large Language Models (LLMs) and Large Reconstruction Models (LRMs), which can generate 3D volumes from a single image or from textual instructions. The methodology presented in this article combines the LLM Google Gemini 2.5 Flash Image, also known as Nano Banana, with commercial LRM-based programs such as Archsynth, Meshy, Hexagen, and Hitem3D. The workflow begins with the generation of a high-fidelity 45° isometric view from historical photographs, which then serves as the basis for producing 3D models with realistic textures and lighting. Several case studies are analysed, ranging from surviving monuments, such as the Tower of the Scipios, used to assess accuracy, to lost heritage sites such as the Portal del Mar, Casa Xifré, Portal del Francolí, Casa Güell, and the public wash house of Vila-seca. The results reveal a critical duality. Although the visualisations are impressive and plausible, GenAI also produces serious errors and lacks geometric reliability. A notable example is Casa Xifré, where the GenAI tool represented 17 arches on the façade instead of the 21 actually present. In the case of the Tower of the Scipios, the algorithm invented a non-existent interior chamber, despite the monument being solid, and duplicated sculptural reliefs. Metric tests carried out with CloudCompare indicate that, while simple geometries may be reproduced with acceptable accuracy, complex geometries and areas inf, Este trabajo fuerealizado bajo el soporte económico del Centro de Estudios sobre Conflictos Sociales (CECOS), Centro de Investigación de la URV,por la Generalitat de Catalunya a través del programa el Suport a la Generació de la Recerca(SGR) del proyecto MIRMED-GIAC [2021 SGR 00790] y con el soporte del Aula Taller Telefónica UB-URV -Smart Cities-,del Proyecto europeo CHANGES 2024-1-CY01-KA220MOD-000256266 Cultural Heritage and Archaeology of study to climate changes in Antiquityy del proyecto Mesjenet, Agencia Estatal de Investigación, AEI PID2021-128069NB-I00.