FreeLayout: Free range multi-modal input to free form floorplan generation

Abstract

Automated floorplan generation has advanced rapidly, yet most methods remain constrained to orthogonal geometry, single-scale synthesis, and fully specified inputs—mismatched to real practice where designers iterate from sparse, evolving cues and irregular floor footprints. We present FreeLayout, a free-form floorplan generation framework that adapts an SDXL-class latent diffusion backbone to structured architectural planning with reliability under partial multimodal guidance. FreeLayout introduces (i) a switchable dual-mode diffusion process that models the floor-to-apartment hierarchy, generating floor-level apartment segmentation and apartment-level room layouts within one unified model; (ii) a Dynamic Masking Partial Input (DMPI) module that fuses incomplete design programs, layout graphs, and freehand sketches into robust text-conditioning embeddings; and (iii) an attention-based grounding regularizer that aligns prompt tokens with spatial regions, improving boundary adherence and enabling interpretable error diagnosis. Unlike prior image generation approaches, FreeLayout embeds layout-graph structure directly into the denoising process to better preserve adjacencies and program semantics, while supporting iterative addition, removal, and editing of free-range constraints. Experiments on SwissDwellings (single-apartment), MSD, and the RPLAN dataset show that our method improves layout coherence, constraint adherence, and robustness over state-of-the-art baselines under both fully specified and sparse inputs.