Masked irregular layout planning from partial input with fused modalities

Floorplan layout planning is time-intensive and requires balancing diverse design constraints. While recent generative models can automate layout synthesis, they are often restricted to rectangular geometries and offer limited support for iterative customization. This paper presents MaskLayout, a graph-structured generative transformer that (1) generates irregular room polygons within non-rectangular site boundaries, (2) supports a wide range of partial user inputs for iterative refinement, (3) jointly models six layout attributes (type, location, size, adjacency, region, and windows), and (4) strengthens room–attribute reasoning via dual vector–image modalities with enhanced cross-attention. A benchmark of more than 43,000 apartment layouts with irregular building contexts was constructed, and MaskLayout was trained using dynamic masking to simulate incomplete user guidance. Parallel decoding was also employed to enable efficient inference. Extensive evaluations show that MaskLayout outperforms state-of-the-art methods across multiple metrics and produces high-quality irregular layouts suitable for practical floorplan design.