GFLAN: Generative Functional Layouts

TL;DR

GFLAN addresses residential floor-plan generation by factorizing the task into topology-focused room-center placement and geometry-focused rectangle regression. It uses a dual-encoder CNN to produce multi-type center heatmaps and a transformer-based GNN to infer interior room boundaries from a hybrid room–boundary graph, incorporating balcony attachments and footprint constraints. The two-stage training, ablations, and diverse outputs demonstrate improved adjacency, circulation, and program realism with efficient inference. The approach achieves more usable, connected layouts than prior methods and offers multiple valid design options for early-stage architectural exploration.

Abstract

Automated floor plan generation lies at the intersection of combinatorial search, geometric constraint satisfaction, and functional design requirements -- a confluence that has historically resisted a unified computational treatment. While recent deep learning approaches have improved the state of the art, they often struggle to capture architectural reasoning: the precedence of topological relationships over geometric instantiation, the propagation of functional constraints through adjacency networks, and the emergence of circulation patterns from local connectivity decisions. To address these fundamental challenges, this paper introduces GFLAN, a generative framework that restructures floor plan synthesis through explicit factorization into topological planning and geometric realization. Given a single exterior boundary and a front-door location, our approach departs from direct pixel-to-pixel or wall-tracing generation in favor of a principled two-stage decomposition. Stage A employs a specialized convolutional architecture with dual encoders -- separating invariant spatial context from evolving layout state -- to sequentially allocate room centroids within the building envelope via discrete probability maps over feasible placements. Stage B constructs a heterogeneous graph linking room nodes to boundary vertices, then applies a Transformer-augmented graph neural network (GNN) that jointly regresses room boundaries.

Figures (21)

• Figure 1: GFLAN—envelope to layout. Given the exterior envelope and front-door location (a), GFLAN generates a functional floor plan (b) that satisfies adjacency, area, and envelope constraints.
• Figure 2: Overview of the two-stage pipeline. A convolutional layout encoder produces latent representations from the building envelope and room program. Sequential decoders then predict room centers conditioned on previously placed rooms. A transformer-based graph neural network regresses full room boundaries while respecting structural constraints.
• Figure 3: Graph structure used in Stage B. Room nodes connect to neighboring room nodes and nearby boundary nodes. The boundary is discretized into corner/edge nodes.
• Figure 4: Qualitative comparison on four envelopes. Graph2Plan often yields trapped or extreme-aspect rooms and unbalanced zoning; WallPlan can bottleneck circulation or misserved bedrooms (e.g., one restroom for three bedrooms). GFLAN keeps all rooms reachable, preserves near-rectangular shapes, and localizes irregularity to boundary fit. (For the effect of counting balconies as interior floor area, see the “GFLAN-B” ablation in Appendix \ref{['app:balconies']}, Fig. \ref{['fig:ablation-balcony']}.)
• Figure 5: Program diversity. GFLAN covers a broader range of restroom and bedroom counts, while Graph2Plan and WallPlan outputs concentrate in a narrow band (e.g., often one restroom).