LCA-GCS: Large City Architecture - Generated Cityscape Set — Lab for Environmental Design Strategies

LCA-GCS: Large City Architecture - Generated Cityscape Set

Dataset
2024
https://www.largecityarchitecture.org

The Large City Architecture - Generated Cityscape dataset (LCA-GCS) is a comprehensive collection of 1,060,166 AI-generated images representing architectural features of 5,856 global cities. Created using advanced diffusion models, this dataset offers over 200 samples of various architectural types for each city with a population exceeding 100,000. LCA-GCS aims to facilitate comparative analysis, synthesis, and learning from AI-represented architecture, providing a unique resource for researchers, urban planners, and AI developers interested in global architectural trends and urban design.

The LCA-GCS-set represents cities by situating architectural features from details to interiors to urban forms uncommon for existing databases in a flat ontology that allows to search and compare architectural representations in various ways. We hope that the LCA-set benefits future research that could identify recurring spatial configurations, material palettes, and design elements characteristic of specific regions, cultures, or building typologies. Building on it, the synthetic dataset can contribute to e.g., training site-specific models, identifying commonalities, or extracting features that are difficult to access in existing data.

You can find a machine-readable version to download here: https://huggingface.co/datasets/Punktiert/LCA-GCS

Citation: Koehler, D. (2024). LCA-GCS: Large City Architecture - Generated Cityscape Set - [Dataset]. Huggingface. https://huggingface.co/datasets/punktiert/LCA-GIS, https://doi.org/10.57967/hf/3111, www.largecityarchitecture.org

Geographic distribution visualization of the Large City Architecture dataset, displaying 15 architectural samples for each of 5,856 global cities (87,840 total images).

Geographic distribution map of a single architectural typology generated with consistent seed values across global locations.

Quantitative analysis of image alignment with the prompt, visualized through CLIPScore metrics (blue indicating low alignment between text prompts and generated images; yellow indicating high alignment), revealing geographic patterns of architectural documentation bias in model training data.

Calgary represented by 100 generated samples within the LCA-GCS dataset.

Generated Images of typical buildings. Left: Antananarivo, right: Quito.

Selection of discrete types within a prompt walk, generated with SD 2.1.

Sequence of a ‘prompt-walk,’ revealing abrupt transitions between building typologies. These sudden formal shifts suggest that diffusion models organize architectural knowledge in distinct conceptual clusters rather than continuous gradients, indicating how the model’s latent space discretizes architectural understanding into separate typological regions rather than smooth continua.