Generative AI has found its way into creative production in a variety of ways. The creation of spatial structures, though, has been largely absent from AI-based generation processes. The main reason for this absence is the lack of training data sets with pairs of 3d-structures and descriptive text. With this project we are exploring possible approaches of generating such pairs. Film scripts and their corresponding film images represent a rich resource for spatial structures and their descriptions. We are investigating methods to leverage this resource by identifying descriptive passages and correlating them with 3-dimensional reconstructions of the corresponding spatial structures.