A Photo to 3D Workflow for Generation of LOD3 Digital Building Representation

Abstract

This study proposes a sequential AI workflow for generating LOD3-oriented digital building representations from a single architectural photograph. The novelty of the approach lies in the use of Gemini Flash 2.5 for preprocessing and Hunyuan3D 2.1 for image-to-3D generation within a preprocessing–reconstruction decoupled pipeline. Rather than directly producing a 3D model from an uncontrolled photograph, the workflow first normalizes the image by isolating the building, converting the scene to daylight, and approximating an isometric view, and then converts this image into a textured 3D mesh. In this way, the method offers a rapid, low-data, and reproducible alternative for cultural heritage visualization, comparative analysis, and digital archiving. Experiments were conducted on 16 photographs of architectural structures representing typologies from the early Republican and Ottoman periods. The findings demonstrate that the method offers a tool for the digital documentation of cultural heritage, archiving and academic research. The resulting 3D outputs are suitable for urban silhouette analysis at the LOD3, volumetric comparison, testing urban design scenarios, and the creation of digital heritage archives. The method works with a single photograph and cloud-based AI models, which is advantageous when on-site scanning is difficult, risky, or costly. Furthermore, the workflow can be enhanced with multiple images, text input, and other sensor data to produce digital twins with geometric accuracy. In this respect, the study demonstrates that by combining cutting-edge technologies in 3D modeling from a single image, 3D twins of architectural objects can be produced, and proposes an AI-powered method for creating digital architectural representations.