GFLAN: Generative Functional Layouts
Mohamed Abouagour, Eleftherios Garyfallidis
TL;DR
GFLAN introduces a two-stage generative framework for floor plan synthesis that explicitly separates topological planning from geometric realization, improving architectural reasoning in automated layout generation.
Contribution
The paper presents GFLAN, a novel generative approach that factorizes floor plan creation into topological and geometric stages, addressing limitations of prior deep learning methods.
Findings
Effective separation of topological and geometric planning.
Improved layout quality over existing methods.
Demonstrated capability to generate feasible floor plans.
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…
Peer Reviews
Decision·ICLR 2026 Conference Desk Rejected Submission
1. **Principled Two-Stage Decomposition:** The explicit factorization of the problem into topological planning (room centers and adjacency) and geometric realization (precise rectangles) is a significant strength. This approach mirrors human architectural thought and allows for the separate handling of combinatorial complexity (topology) and geometric constraint satisfaction. 2. **Structurally and Functionally Aware Output:** The GNN-based geometric realization stage is designed to ensure struc
1. **Limited Empirical Validation Scope:** The paper's empirical evaluation is constrained to a single benchmark dataset and comparison against only two key baselines (one being a prior work from the same lab). While the results on this benchmark are strong, the efficiency and generalizability of GFLAN would be more robustly proven by including a broader set of established generative floor plan benchmarks and comparing against a wider array of modern architectural generation techniques. 2. **Or
1. The paper propose a new factorization considering both topology and geometry, based on hybrid room–boundary graph and boundary-aware clipping. It contributes to the performance of the method. 2. It provides clear and detailed method introduction including detailed architectural description (dual-encoder CNN; TransformerConv GNN) and training schedule 3. The paper improves human-aligned functional criteria and circulation characteristics over SOTA baselines.
1. The proposed method only compared on very limited metrics (only functionality metrics) to very limited amount of baselines (2 baselines from 2020 and 2022). 2. The sequential generation of the centriod may suffer from the lack of randomness. How do you evaluate it? 3. Why the two stage structure better than the end-to-end other methods (either DM or transformer based). There is lack of discussions. For example, how sensitive is performance to errors in the program-count predictor?
1) A clean topology-then-geometry factorization: sequential room-center heatmaps with a dual-encoder CNN, followed by a TransformerConv GNN on a hybrid room–boundary graph that regresses rectangles with footprint-aware clipping. This design targets adjacency/circulation intent before geometry, reducing failure modes of prior pipelines. 2) The staged interface is easy to follow and editable: users can adjust/add centers after Stage A and let Stage B realize geometry, which make the model more fle
1) The method targets single-story, single-envelope, axis-aligned residential layouts and omits structural/HVAC constraints—reducing external validity for multi-story or non-orthogonal plans and non-residential programs, while mentioned as limitation by authors, but it hinders the practicality of real application. 2) The model is trained on the ResPlan dataset but evaluated on the RPLAN test set. While the authors frame this as a test of generalization, this cross-dataset evaluation could introd
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
Topics3D Shape Modeling and Analysis · Computational Geometry and Mesh Generation · 3D Modeling in Geospatial Applications