FZplan: Automated Space Layout Planning Dataset Generation Using Fuzzy Maps

Training data-driven generative models for architectural layout generation requires large-scale datasets with continuous, multi-dimensional quality annotations, yet existing floor-plan corpora rely on binary validity checks and offer limited modality coverage. This paper presents FZplan, an automated pipeline that synthesises validated residential floor plans by coupling topology enumeration with Mamdani-type fuzzy inference—embedding architectural quality evaluation within the generative loop rather than applying it post-hoc. The pipeline follows four stages: circulation synthesis, topology enumeration, fuzzy-guided spatial placement, and six-check hard-constraint validation. At the core is a fuzzy inference system that scores every candidate room placement on five spatial quality dimensions—proximity, adjacency, privacy, view, and accessibility—using 32 expert-defined rules implemented in the system. A precomputed five-dimensional lookup table (215 ≈ 4 million entries) with five-dimensional multilinear interpolation over the 32 vertices of each local hypercube accelerates fuzzy evaluation by ≈500×, reducing per-placement cost from ∼5 ms to ∼10 μs and enabling real-time guidance. Applied to five residential typologies (studio through four-bedroom), the pipeline achieves 100% hard-constraint pass rates and produces 21,379 validated plans with stable mean composite quality scores in the range 0.819–0.835 across typologies. A symmetry augmentation module scales this corpus to over 170,000 synchronised records across seven modalities (images, ControlNet maps, graphs, grids, prompts, statistics, and metadata). The resulting dataset embeds continuous fuzzy quality scores at the record level, providing a transparent, reproducible, and ready-to-use resource for training and benchmarking downstream generative architectural models.