What is the role of graph neural networks (GNN) in the cross-modal spatial layout generation method proposed in this chapter?

In the proposed method, the GNN acts as the generative model that receives feature embeddings created by AI agents from text, graph, and boundary inputs, and outputs node and edge features that are converted into a detailed 3D spatial layout. Nodes represent spatial grids and edges represent wall categories, so the GNN learns to infer layout details such as wall types and spatial divisions from the fused cross-modal inputs.

The method is structured into three parts: input representation, agent-driven feature engineering, and the GNN generation model. After graph, text, and boundary inputs are unified into node and edge features, these features are fed into a trained GNN model. The GNN predicts node and edge features, which are then transformed into a new three-dimensional room layout. The graph encoding used by the GNN represents nodes as spatial grids and edges as wall categories, with node features including coordinates, boundary types, and room categories. The model is trained by comparing predicted edge categories with ground-truth edge types using cross-entropy loss, allowing the GNN to learn spatial layout logic. The GNN is also described as effective at integrating boundary constraints with topological constraints, making it suitable for generating vector models that can be adjusted through textual input, anchor points, and boundaries.