Latent Earth

I asked one image model, FLUX.2, to draw 40,000 real places on Earth, giving it nothing but each place’s name: no photographs, no maps, nothing to correct it. Each answer was then read back by a vision-language model and rendered a second time from its own description, four times per place. The result is not a picture of the world but a probe of the model: a survey of what a large model believes architecture looks like, place by place.

The atlas sorts 160,000 of these images by the model’s own internal representations, recorded while each image was being generated. No clustering, no regions drawn: whatever order appears on the wall emerged from the sort. Places the model knows well spread into distinct neighbourhoods of the wall; places it has barely seen collapse into a few recurring clichés. That compression is the finding, and the atlas is built so that it stays visible.

The full corpus is public. Every tile carries its place name, its prompts, and five recorded internal representations, so the measurement can be repeated, extended, or turned against the next model. The wall itself, The City as a Generated Building Atlas, was commissioned by the Deutsches Architekturmuseum in 2026 for its permanent collection and is in preparation.