Text-conditioned irregular floorplan generation via composable inpainting diffusion

Floorplan layout planning is an iterative process, where designers refine spatial configurations from natural-language descriptions and sketches. Most existing methods do not jointly use natural language and layout graphs as hybrid guidance and are largely limited to rectangular layouts, making it difficult to handle irregular geometries common in practice. This paper presents TextPLAN, a text-conditioned diffusion framework for generating irregular floorplans under flexible multimodal inputs. Users can specify room types, sizes, locations, and spatial relations through free-form text, layout graphs, and on-canvas sketches, where text is converted into structured prompts by a fine-tuned LLM and injected into the diffusion backbone. The framework combines (1) an irregular-shape generation pipeline that respects site boundaries; (2) composable diffusion modules that align sparse structured text with spatial regions; and (3) an inpainting mechanism that restricts generation to a boundary mask. On the SwissDwellings-based benchmark, TextPLAN outperforms strong baselines in visual quality and adherence to user constraints.