Abstract
Generating realistic, large-scale urban layouts remains a significant challenge in computational urban design. Previous deep learning approaches often generate city blocks in isolation, leading to disjointed or repetitive patterns that lack the stylistic consistency of real-world cities. We present COHO, a novel graph-based framework that treats the city as a hierarchical graph. By employing a Graph-based Masked Autoencoder (GMAE) and a priority-based iterative sampling strategy, COHO ensures context-sensitivity, where the layout of each block harmonizes with its neighbors. Evaluated on a massive new dataset of 330 US cities, COHO demonstrates superior performance in realism and urban harmony, significantly outperforming existing state-of-the-art models.
Introduction
The creation of realistic city layouts is essential for applications ranging from urban planning to autonomous vehicle simulation. However, existing methods often struggle with the scale and complexity of urban environments. Procedural generation relies on rigid, hand-coded rules, while many deep learning models fail to capture the "context" of a block—how it relates to its neighbors. This results in "patchwork" cities where blocks appear either overly similar or completely unrelated.
COHO addresses this gap by modeling the city as a hierarchical graph, allowing the generation process to be informed by the surrounding urban fabric.
Methodology
The COHO pipeline consists of three main components:
- Canonical Graph Representation: The city is represented as a graph where nodes are blocks and edges denote adjacency. This avoids the memory overhead of pixel-based representations and naturally handles irregular block shapes.
- Block-level Variational Autoencoder (BVAE): To manage building complexity, a BVAE with a Graph Attention Network (GAT) backbone quantizes 2.5D building geometries into a discrete "codebook" of tokens.
- Graph-based Masked Autoencoder (GMAE): The core of the system, the GMAE, learns to predict masked block layouts based on unmasked neighbors and the road network.
- Priority-based Iterative Sampling: During inference, the model generates blocks based on confidence levels, using early "anchor" blocks to guide the style of subsequent generations.
The COHO Dataset
To support this research, we introduced the COHO dataset, a comprehensive vector-based benchmark. It covers 330 US cities with populations over 100,000, encompassing over 833,000 blocks and 17.6 million buildings. Data was integrated from OpenStreetMap, Microsoft Building Footprints, and the TIGER road network.
Results and Discussion
COHO achieves a state-of-the-art FID of 23.63, a significant improvement over previous models like GlobalMapper (49.55). A key innovation is the Context Score (CTS), which measures urban harmony. COHO's CTS of 0.21 closely matches real-world data (0.0), whereas competing models exhibit extreme over-diversity or over-similarity.
Conclusion
COHO represents a leap forward in city-scale generation by prioritizing context-sensitivity through a hierarchical graph approach. While challenges remain in modeling highly irregular geometries and functional road-building synergies, the framework provides a robust foundation for future 3D urban modeling and high-fidelity simulations.
Summary updated