How does the RoomGNN approach convert bubble diagrams and building boundaries into a format suitable for AI generative models?

RoomGNN first uses AI agents to convert bubble diagrams and building boundaries into structured text. Those structured data are then turned into room feature embeddings by mapping room coordinates inside the boundary, applying a Signed Distance Function to color potential diffusion areas, and overlaying a grid to compute influence probabilities. These embeddings become node features for the GNN, whose edges represent wall categories, enabling the model to generate spatial layouts.

The conversion is carried out through LLM-based agent-driven feature engineering. A 'bubble diagram to text' agent reads bubble diagrams that contain room coordinates and areas and writes them as text, while a 'building boundary to text' agent translates building boundary vectors into structured text. Based on these structured text representations, a room-feature process maps the coordinates of different room types within the building boundary, uses the Signed Distance Function (SDF) to color the likely diffusion areas of functional rooms, overlays a grid over the diffusion areas for evaluation, and calculates the influence probability of graph and text inputs on the properties of the final layout nodes. This produces unified feature embeddings that are passed as input to the generative model. In RoomGNN, the layout is encoded as a graph in which nodes represent spatial grids with features such as coordinates, boundary types, and room categories, and edges represent walls encoded using a one-hot scheme for wall types. After training, the feature embedding derived from user input is fed into the trained GNN, which predicts node and edge features; those features are then transformed into a 3D room layout.