How do House-GAN++ and the method by Zheng and Petzold differ in their approach to generating floor plan layouts?

House-GAN++ directly represents a floor plan as a graph, where rooms are nodes and their relationships are edges, and it generates the layout conditioned on room count, types, and areas. Zheng and Petzold differ by separating the prediction of room topology from the prediction of room geometry, using subgraph neural networks for topology and neural-guided plan sketching for geometry. In short, House-GAN++ uses a unified graph-based generative process, while Zheng and Petzold decouple topological reasoning from geometric layout generation.

According to the related works section, House-GAN++ is an intelligent framework that represents room layouts as graphs. In this representation, nodes correspond to rooms and edges correspond to relationships between rooms. The framework then generates layouts based on input conditions such as the number of rooms, room types, and room areas. This makes the generation process directly tied to a single graph-based model. Zheng and Petzold's method, in contrast, explicitly decouples topological and geometric predictions. They use subgraph neural networks to predict room topologies, and then combine these predictions with neural-guided plan sketching to produce layouts that have reasonable topological structures. Thus, the core difference is that House-GAN++ generates geometry from the graph representation in an integrated way, while Zheng and Petzold separate the task into topological prediction and geometric sketching steps.