GRAPH2PIX: A Generative Model for Converting Room Adjacency Relationships into Layout Images

Abstract

With the advancement of mathematics and computer science, deep learning-based generative design for floorplans has increasingly garnered attention among researchers. This study proposes a graph-based deep learning model, graph2pix (G2P) to synthesize floorplans guided by user-defined constraints. By incorporating room area and type information into the nodes of the graph, G2P can generate floorplans tailored to specific user requirements. It contains three submodels: the Translator, Generator, and Discriminator. The Translator serves as the foundational layer, interpreting and mapping room information into a coherent building boundary. Following this, the Generator takes the helm, synthesizing this information to formulate a preliminary floorplan layout. This layout is further refined and evaluated by the Discriminator, ensuring that the final output maintains a high degree of fidelity both to the user's constraints and to architectural feasibility. Our empirical investigation, utilizing metrics such as "area error" and "adjacency error", underscores the model's exceptional ability to achieve user tasks with a high degree of accuracy and efficiency. These findings underscore the potential of G2P as a transformative tool in the domain of automated architectural design.