AI Learns to Design Parks by Thinking in Graphs, Not Pixels

Generative artificial intelligence has already transformed how architects and designers sketch ideas, churning out photorealistic images from simple text prompts. But when it comes to designing something as deceptively complex as an urban park, the technology has struggled. A new study published in Machine Learning with Applications proposes a solution that mirrors how human designers actually think: before the AI draws anything, it first reasons about the park as a network of interconnected spaces. The result is a hybrid model that combines graph neural networks with generative adversarial networks, producing park layouts that are not only visually convincing but structurally sensible.

The problem with existing approaches, according to the research team led by Yuhan Zhang and Jing Zhao, is that conventional image generators operate at the pixel level. They learn what parks look like but not how parks work. Real park design is a decision-making process in which functional spaces act as nodes and spatial relationships act as edges in a network. Designers must balance where to place plazas, sports areas, and ecological zones, how those zones connect to one another, and how pathways, water bodies, and vegetation knit everything together. Standard generative models have no built-in mechanism to enforce that spatial logic, which is why their outputs often contain inconsistencies, such as activity areas stranded far from any path or circulation routes that make no functional sense.

The researchers’ framework works in two sequential stages. In the first, a GraphSAGE model learns the semantic relationships and spatial connections among park functional nodes. Users provide a site condition image and a set of functional requirements, and the model predicts which pairs of functional zones should be connected, generating an explicit relationship graph. Crucially, the team defined park relationships more broadly than simple geometric contact: two zones count as connected if they share a boundary, if they are linked through transitional elements such as roads, plazas, or water bodies, or if they form part of a continuous circulation route from entrances to core public spaces to secondary activity areas. This composite definition captures the layered, dynamic nature of park spatial organization that wall-based architectural models cannot.

In the second stage, the predicted relationship graph is transformed into a spatially structured functional zoning diagram using park design standards and a Voronoi post-processing algorithm, which allocates pixel areas within the irregular site boundary according to each node’s target size. A CycleGAN model then maps that zoning diagram onto a concrete park design scheme, refining visual details and spatial morphology. The team added two custom loss terms, a region semantic consistency loss and an instance semantic consistency loss, to the standard CycleGAN objective, explicitly forcing the generator to preserve regional correspondence among roads, green spaces, and water bodies and to maintain the boundaries of independent functional blocks.

Building the training data was a substantial undertaking in itself. The researchers collected 9,286 urban park design scheme drawings from professional design repositories and online platforms worldwide, focusing on medium-scale comprehensive parks of 10 to 20 hectares, a range chosen because such parks serve district-level populations with complete composite functions while exhibiting similar design logic in spatial granularity and composition. From these, 198 schemes with clearly delineated functional zoning were manually screened and annotated using LabelMe, then augmented to 1,188 graph samples. Functional spaces were consolidated under a blue-green-grey infrastructure classification framework into a three-tier system spanning plazas, sports and fitness areas, ecological green spaces, waterfront spaces, and other major types.

The experiments showed that the graph generation stage performs robustly. As the training set expanded from roughly 100 to 800 samples, the model’s test AUC climbed from 0.8207 to 0.8991, with the steepest gains occurring below 400 samples and performance plateauing beyond 600. When compared head-to-head against GCN, GAT, and GIN architectures on identical data, GraphSAGE achieved the best results, with a validation AUC of 0.909 and an accuracy of 0.846, a finding the researchers attribute to its inductive neighborhood sampling, which suits park graphs where node types and quantities vary and adjacency relationships are sparse. The generated graphs also reflected genuine design principles: open lawns, water bodies, and plazas emerged as highly connected central hubs, while dense forest areas sat at the periphery as ecological buffers, exactly as human designers would organize them.

The full hybrid model’s most striking advantage appeared in planning utility metrics. Compared with a CycleGAN baseline and Stable Diffusion, the full model achieved a circulation connectivity index of 70.2 percent versus 46.7 percent for the baseline and 38.2 percent for Stable Diffusion, along with a functional adjacency compliance rate of 78.4 percent and a node accessibility ratio of 82.5 percent. It also recorded the lowest Fréchet Inception Distance at 0.3190, indicating the closest match to the real distribution of park layouts. Interestingly, the baseline CycleGAN actually scored higher on pixel-level similarity metrics such as PSNR and SSIM, suggesting that the graph constraints deliberately shift the model away from pixel-perfect replication and toward diverse layouts that obey functional logic.

Ablation experiments confirmed that the graph reasoning component carries most of the weight. Removing GraphSAGE caused the largest performance degradation, with FID rising from 0.3190 to 0.3634, while removing the rule-based coordinate assignment or Voronoi allocation produced smaller declines, indicating those components mainly handle geometric translation rather than functional reasoning. Semantic-category analysis showed the largest FID improvements for roads, water bodies, and green space, the elements with strong spatial organization characteristics, while buildings, which occupy small and sparse areas, saw little benefit. A panel of 12 landscape professionals scored the hybrid model’s outputs above 7 out of 10 on all six evaluation dimensions, including spatial structure, park roads, and usability, consistently higher than the baseline, though still below manually designed ground truth.

The researchers are careful to position the tool as an assistant rather than a replacement. The framework is designed for the early conceptual phase of design, where it can rapidly generate layout drafts with basic spatial logic in about 2.12 seconds per 512-by-512 image, allowing designers to explore and compare schemes before committing to detailed development. It embeds the generative model within the knowledge-driven workflow that designers already follow, from requirement analysis through functional zoning to spatial composition, rather than attempting to bypass that reasoning with a purely image-to-image translation.

Limitations remain. The training data concentrates on medium-scale comprehensive parks, so the model is not yet suited to pocket parks or very large green space systems, and the unsupervised CycleGAN stage still produces limited detail in small-scale elements and textures. The team suggests future work could expand the dataset across park types and styles and enrich the graph inputs with terrain elevation, surrounding context, and existing vegetation. But the core insight stands: teaching generative AI to reason about space as a graph before it renders an image produces park designs that are more rational, more usable, and far closer to what professional landscape architects would actually build.

Subject of Research: Graph-constrained generative AI for automated urban park layout design

Article Title: Graph-constrained generative adversarial network for automated urban green space layout generation

Article References: Zhang, Y., Liu, Y., Ni, B., Chen, R., & Zhao, J. (2026). Graph-constrained generative adversarial network for automated urban green space layout generation. Machine Learning with Applications, 26, Article 101020. https://doi.org/10.1016/j.mlwa.2026.101020

Image Credits: AI Generated

DOI: Not provided

Keywords: generative adversarial networks, graph neural networks, urban parks, landscape architecture, CycleGAN, GraphSAGE, generative design, spatial layout generation, urban green space, artificial intelligence, Voronoi diagram, human-AI collaboration