Early Rule-Based Approaches (1970s–1980s)
The quest to automate floor plan design began as early as the 1970s. One of the first conceptual milestones was the introduction of shape grammars by Stiny and Gips in 1971 – a formal rule-based system for generating designs, including architectural layouts, via recursive shape transformation rules. These early approaches were largely symbolic or rule-driven: for example, Friedman’s 1971 work attempted algorithmic space planning, and Mitchell’s 1976 theory laid out a method to enumerate possible room arrangements. By the late 1970s and 1980s, researchers were exploring expert systems and heuristic rules for the “space allocation problem.” Shaviv’s work in 1974 and 1987 on computerized space allocation exemplified the era’s logic-driven methods for layout planning.
Optimization and Evolutionary Methods (1990s–2000s)
In the 1990s, AI in architecture shifted towards search and optimization algorithms to handle the combinatorial complexity of layouts. Pioneering work by John Gero and colleagues introduced evolutionary algorithms for floor plan design: Jo and Gero (1996) simulated simple architectural plans using genetic algorithms, and Rosenman et al. (1997) extended this by combining genetic algorithms with genetic programming. Throughout the late 1990s, multiple researchers applied evolutionary strategies to generate spatial configurations.
By the early 2000s, the field saw a proliferation of heuristic methods like simulated annealing and hybrid methods that combined evolutionary algorithms with constraint programming or shape grammars. A notable trend was the use of multi-objective optimization to balance functional requirements with performance criteria.
Data-Driven Paradigms and Key Turning Points (2010–2015)
A pivotal shift came with the 2010 paper by Paul Merrell et al., titled “Computer-Generated Residential Building Layouts.” It introduced a Bayesian network trained on real-world floor plans to learn room connectivity patterns, and a stochastic optimization step for layout geometry. This work marked the beginning of referential methods that learn from prior design data.
Following this, early machine learning efforts emerged to analyze and label architectural drawings. These were transitional years where traditional rule-based methods coexisted with data-driven modeling, awaiting larger datasets and more powerful learning frameworks.
Deep Generative Models Revolution (2016–2020)
With deep learning, automated floor plan generation saw significant progress. Large datasets like RPLAN (2019) and LIFULL Home Dataset enabled training of deep neural networks. Initial CNN-based methods struggled with multi-room consistency, but Generative Adversarial Networks (GANs) changed the landscape.
- House-GAN (2020): Introduced a graph-constrained GAN to generate layouts from bubble diagrams using graph neural networks.
- Graph2Plan, FloorplanGAN, and House-GAN++ further improved realism and layout coherence.
- Models could now generate both rasterized floorplans and precise vector-based layouts with high fidelity.
Conditional generation also advanced, allowing layout generation within fixed boundaries or based on room adjacency constraints.
Recent Advances: Graph Networks, Multi-Story and Diffusion Models (2021–2025)
Recent models target multi-floor generation and layout refinement using advanced architectures:
- Building-GAN (2022) and Building-GNN (2023): Generate stacked floor plans using graph neural networks.
- HouseDiffusion (2023): Applies denoising diffusion to generate room polygons from bubble diagrams.
Newer models integrate building codes and real-world constraints into layout generation. Diffusion models improve stability over GANs, and evaluation benchmarks are more standardized.
There is also increasing focus on text-conditioned generation and integration with large language models (LLMs) for brief interpretation and co-design.
Toward the Future
The development of AI for floor plan generation evolved from symbolic systems to deep learning models capable of generating viable architectural layouts from simple inputs.
Key turning points:
- 1971: Shape grammars
- 1990s: Genetic and optimization algorithms
- 2010: Merrell’s Bayesian layout synthesis
- 2020s: Deep GANs, diffusion models, and graph networks
The field continues to move toward intelligent systems that understand human design intent and produce functionally valid, creative architectural plans. AI is rapidly becoming a collaborative partner in architectural design.