Punktiert/lca-gcs-explorer

LCA-GCS explorer

A browser for the Large City Architecture: Generated Cityscape Set: 1,060,166 images of 5,856 cities across 86 architectural types, generated with Stable Diffusion 2.1, SDXL 1.0, SDXL Turbo and SD 3 Medium (Daniel Koehler, The University of Texas at Austin, 2022 to 2024; CC BY-NC 4.0; DOI 10.57967/hf/3111). It replaces the earlier site largecityarchitecture.org and runs as well on lab-eds.org.

• —Cities: search a city, filter by country, sort by population or CLIP score.
• —City: every image of one city in one grid per model, ordered by type, with the prompt and seed in the lightbox; keyboard paging; a link to the shard the image sits in.
• —Compare: one type across every city, from the largest city down, more rows as you scroll.
• —Maps: the maps assembled from one generated image per place; zoom with the wheel, drag to pan.
• —CLIP: what the score means, the most and least recognisable cities with their thumbnails, the distribution, the countries, and the world map.

How it works

One index.html, no build, no server. The index (viewer/index/), the thumbnail sheets (viewer/thumbs/), the maps and the country outlines are small files in the dataset repository. The images themselves are not copied anywhere: the explorer resolves a shard (dataset_chunk_NNNNN.tar) once per session and reads each image out of it by HTTP byte range, straight from the dataset's storage. A full city view touches about 140 shards and 180 range reads.

?base=<url> points the explorer at another copy of viewer/ (for local testing: python tools/lca/serve.py <folder> 8765 serves one with CORS and range support). The index is built by tools/lca/index_tars.py (shard offsets), tools/lca/build_index.py (cities, types, prompts, coordinates, CLIP scores) and tools/lca/make_sheets.py (thumbnail sheets, maps, outlines).

Citation

Koehler, D. (2024). LCA-GCS: Large City Architecture, Generated Cityscape Set. Hugging Face. https://doi.org/10.57967/hf/3111