Towards Generative Twins: An AI Based Pipeline from Explicit 3D Reconstruction to Probabilistic Synthesis

Deploying digital twins across the broader building stock is constrained by a key bottleneck: high-quality 3D content from LiDAR or photogrammetry remains labor-intensive and difficult to scale. We introduce Generative Twins (GT), a workflow that shifts digital twin creation from explicit geometric reconstruction toward probabilistic 3D synthesis. The pipeline takes a single street-level façade photograph, corrects perspective distortion and removes occlusions, uses an instruction-conditioned multimodal image model to synthesise an isometric view of the building, and passes that view to an image-to-3D reconstruction model that leverages learned shape priors to produce a textured mesh without manual modelling. Three image to-3D platforms (Hitem3D, Hyper3D and Hunyuan3D 3.1) are compared on identical isometric inputs and assessed on architectural boundary clarity. We demonstrate GT through two applied cases: (i) reconstruction of ordinary residential streets in Detmold as stimuli for a virtual-reality study of multi sensory urban perception, and (ii) Green Editor, a design-exploration tool in which façade greenery is added or removed at the image layer before mesh generation. Together these cases position Generative Twins as a scalable pathway for the digitization of ordinary building stock, with emphasis on controllability, editability, and iterative design workflows.