Abstract: We introduce a deep learning system for automatic floorplan generation from layout graphs. Our system combines generative modeling with user-in-the-loop design in which users can add sparse constraints like room numbers, connectivity, and other layout adjustments. The system relies on a core deep neural network that takes an input building boundary and layout graph to generate realistic and constraint-abiding floorplans. The system utilizes a graph neural network (GNN) to encode layout patterns and convolutional neural networks (CNNs) for processing building contours and rasterized floorplan images. The model, trained on RPLAN, a 80K-annotated floorplan dataset, outputs varied floorplan layouts consistent with user inputs. We measure its performance via qualitative and quantitative analysis, ablation experiments, and comparison against state-of-the-art techniques, proving its effectiveness and flexibility in floorplan synthesis with constraints.

Keywords-floorplan generation, layout graph, deep generative modeling

| DOI: 10.17148/IARJSET.2025.12227