Abstract
BlockPlanner is a novel generative framework designed for synthesizing realistic city blocks using a vectorized graph representation. By modeling land lots as nodes in a "ring topology" within a Variational Autoencoder (VAE) framework, the system generates diverse and valid urban layouts. This approach significantly outperforms traditional procedural or pixel-based methods in both geometric precision and computational efficiency.
Introduction
City modeling is a cornerstone of urban planning, navigation, and digital environment design. Historically, this has relied on procedural modeling—rigid, hand-coded heuristics that often fail to capture the organic complexity of real-world urban fabrics. Recent attempts to adapt indoor floorplan generation techniques to city blocks have faced significant challenges. Rasterized (pixel-based) approaches struggle with the high resolution required to capture small land lots, leading to high computational costs and imprecise, "fuzzy" geometries.
BlockPlanner addresses these limitations by introducing a representation that is both precise (vectorized) and aware of the global structure of the block.
Methodology: The Ring Topology and Graph VAE
The core innovation of BlockPlanner is treating a city block as a two-tier graph. Instead of pixels, each land lot is represented as a vector containing its 3D geometry (centroid, dimensions, building height) and land-use semantics (e.g., residential, commercial).
To ensure every lot is accessible from a street—a fundamental urban regulation—the authors arrange land lots into a ring topology that follows the block's boundaries. This structure reduces data variance and allows the model to learn intrinsic "co-boundary" relationships effectively.
The framework utilizes a Graph VAE:
- Encoder: Uses message passing to aggregate local lot features into a global block-level latent code.
- Decoder: Predicts the block's aspect ratio and an adjacency "edge map," then uses iterative message passing to refine individual lot geometries.
- Geometric Loss Functions: Specific losses penalize overlapping lots, gaps between neighbors, and misalignments with block boundaries, ensuring the generated blocks are structurally valid.
Data and Evaluation
The model was trained on the NYC-Block Dataset, a large-scale collection of 28,838 city blocks curated from official New York City urban planning resources. Unlike previous datasets, this includes accurate lot geometries and 3D building envelopes.
In comparative evaluations against baselines like BlockGAN and LayoutVAE, BlockPlanner demonstrated superior performance:
- Visual Quality: It achieved a Realism score of 95.6, far exceeding BlockGAN's 63.4.
- Diversity: It recorded the lowest FID score (36.9), indicating high similarity to real-world NYC block distributions.
- Efficiency: By operating on vectors rather than high-resolution images, BlockPlanner uses approximately 1000x fewer FLOPS during inference than pixel-based GANs.
Conclusion
BlockPlanner provides a robust, efficient solution for urban synthesis by prioritizing topology over pixels. While currently focused on rectangular blocks, its success in topology refinement and footprint generation paves the way for more complex urban modeling. Future work will explore irregular "super-blocks" and the integration of these models into large-scale neighborhood synthesis tools.
Summary updated