Punktiert/Latent-Earth

Latent Earth: An Atlas of Architecture in Flux.2

200,000 images of 40,000 places on Earth, each rendered by a single image model in a single state of its training, with five internal representations recorded for every image while it was being generated.

Nothing else enters. Each prompt contains only a place's name; no photographs, no maps, no climate records correct what the model proposes. This is therefore not a depiction of the world but a probe of the model: a survey of what a compression of the internet believes architecture looks like, place by place.

The corpus is the material behind an atlas commissioned by the Deutsches Architekturmuseum (DAM), Frankfurt, and behind the accompanying paper. It is released so the measurement can be repeated, extended, or turned against the next model.

_{The exhibited sort (s_highdim), reduced. Larger version: preview/atlas_s_highdim_wall.jpg.}

Explore the atlas interactively — the exhibited sort, zoomable from the full wall down to individual tiles, each labelled with its place.

---

How every image was produced

Every image is produced twice over. A first pass renders the bare name at low resolution; a vision-language model describes what appeared, in architectural terms; a second pass renders name and description together at full resolution. The model reads its own guess and elaborates it, an inductive loop closed to outside correction. Four seeds run this loop four times for every place, and a fifth pass repeats it with an independent first stage.

1. Step 1 — the generator receives the place name and nothing else, at 512 px. These images are published in images/gen1/.
2. Describe — Qwen2.5-VL-7B-Instruct lists 8–15 visible architectural features as short phrases. A denylist strips style labels ("colonial", "modernist"), place and country names, and narrative vocabulary, so only concrete, transferable descriptions survive.
3. Step 2 — the surviving features return as the prompt, with the place name appended last, at 768 px. Published in images/gen2/.

Setting	Value
Generator	FLUX.2 [dev] (Black Forest Labs), 4-bit quantization (bitsandbytes NF4)
Step-2 resolution / steps / guidance	768 px / 14 / 4.0
Step-1 resolution	512 px
Describing model	Qwen2.5-VL-7B-Instruct
Seeds	blake2s(place_id : sample_idx : phase), 63-bit — deterministic
Completed	199,997 of 200,000 scheduled images with full recordings

Two prompt revisions are present, and this matters for reproduction. 76% of images use the grammar above ("<features>. <name>.", varying only in seed across the four attempts). The remaining 24%, uniformly distributed across attempt slots, additionally prefix a scale-framing phrase that varies by attempt, from close-up roof detail to a whole city building, with correspondingly capped feature counts. step2_prompt always records what was actually sent, so the split is recoverable per row.

What was recorded during generation

During generation each image leaves behind its internal traces: the residual stream of the generator itself, two independent vision encoders, the text encoder's reading of the prompt, and the describing model's reading of the image. All five are captured in-pass, without ever reloading an image, and all five are published.

Channel	Dim	What was captured	Reads as
dino	1024	DINOv2-large CLS token of the finished image	visual appearance, self-supervised
siglip	1152	SigLIP-2 So400m pooled output of the finished image	depicted content, language-aligned
flux_residual	6144	FLUX.2 residual stream, final block, mean over the last denoising step	the generator's own state while drawing
t5_prompt	15360	prompt encoder embedding of the Step-2 request, masked mean over tokens	what the model was asked
vlm_image	3584	Qwen2.5-VL visual encoding of the Step-1 image	how the describing model saw the first draft

flux_residual is the one that matters most for auditing: it is recorded inside the generator while the image forms, so results built on it cannot be explained by the prompt text or by any external encoder.

Vectors are stored as float16, the dtype they were captured in. Patch grids (vae_patches, siglip_patches) and the DAAM attention maps (attn_place, attn_features, all zero — that capture was deferred) are not published as Parquet.

---

Quick start

The dataset is designed to be queryable before it is downloaded. All metadata is Parquet at the root.

1. Look without downloading (a few MB):

import duckdb
q = ("SELECT country, count(*) n FROM "
     "'hf://datasets/Punktiert/Latent-Earth/data/places.parquet' "
     "GROUP BY 1 ORDER BY n DESC LIMIT 10")
print(duckdb.sql(q))

2. Try the lite subset (2,000 places, images and embeddings, a few GB) before committing to the full corpus.

3. Stream the images. Every tar member carries a .json sidecar with its full metadata, so no key parsing is ever required:

from datasets import load_dataset
ds = load_dataset("Punktiert/Latent-Earth", "images", streaming=True, split="train")
ex = next(iter(ds))
print(ex["json"]["name"], ex["json"]["country"], ex["json"]["step2_prompt"])

4. Load embeddings:

import polars as pl
df = pl.read_parquet("hf://datasets/Punktiert/Latent-Earth/embeddings/dino/*.parquet")
X = df["dino"].to_numpy().astype("float32")   # (n, 1024)

Join anything to anything on place_id (a place) or tile_id (one image).

---

Structure

data/
  places.parquet              40,000 rows — one per place, with covariates
  tiles/tiles-*.parquet       200,000 rows — one per image, self-sufficient
  captions.parquet            the describing model's feature lists
  step1.parquet               index of the Step-1 probe images
images/
  gen2/gen2-*.tar             WebDataset: <tile_id>.jpg + <tile_id>.json
  gen1/gen1-*.tar             Step-1 probes: <place_id>__<slot>.jpg + .json
embeddings/
  <channel>/<channel>-*.parquet   tile_id + fixed-size float16 vector
results/                      the paper's analysis outputs (JSON/Parquet)
preview/                      reduced image of the exhibited sort
viewer/                       tile pyramid + labels behind the explorer
lite/                         2,000-place taster: everything, small

Keys

Key	Form	Meaning
place_id	UUIDv5 string	one real place; stable across the whole release
tile_id	<place_id>__<sample_idx>	one generated image
sample_idx	0–4	which pass (0–3 share a Step-1 image; 4 is independent)

tile_id contains no spaces and no delimiter that occurs inside a field, so it is safe as a filename, a tar key and a join column.

data/places.parquet

Column	Type	Meaning
place_id	str	canonical key (UUIDv5 of source, id and 6-dp coordinates)
name	str	the place name — the only field that ever entered a prompt
country, iso2, iso3, admin_name	str	from the source gazetteer
lat, lng	f64	decimal degrees, WGS84
population, population_band	i64 / str	10,000 to 37.7 M (median 24,910)
un_region	str	UN M49 grouping
koppen_class	str	Köppen–Geiger class, point-sampled from the Beck et al. 1-km raster (1991–2020), 5-km spiral fallback over water
wiki_visibility_score	f64	documentation index: PC1 of z-scored [has article, log bytes, log sitelinks, log Commons files]
wiki_visibility_bin	cat	absent / sparse / medium / rich

data/tiles/*.parquet

Place covariates are denormalized onto every row, so the common case needs no join. Adds step2_prompt (the exact string fed to the generator), seed_step1 / seed_step2, n_features, sparse_features, per-stage status flags and image_key.

images/gen1/ — the Step-1 probes

These are the most direct evidence in the dataset: what the model draws from a place name and nothing else, before any feature list exists. Everything downstream derives from them.

They are keyed by place, not by tile — samples 0–3 share one Step-1 image and sample 4 has its own — so the tar keys are <place_id>__<slot>.jpg and data/step1.parquet indexes them.

---

The atlas

The arrangement is the argument. Concatenated and whitened, the recorded traces give every tile a coordinate in a high-dimensional space, and a sorting algorithm places the tiles so that neighbours on the wall are neighbours in that space. Nothing more is done: no reduction to two dimensions, no clustering, no regions drawn, no zones reserved. Whatever large-scale order appears has emerged from the sort.

Read at two distances. At arm's length each tile is one centimetre and adjacent tiles are near twins: a roofline recurs, shifts, thickens; typology morphs step by step. From across the room the field drifts without borders. The space of architecture inside this model is a continuum, dense with recurring types yet nowhere divided into kinds. Where a geographic atlas separates land from water, this one has no coastlines to draw.

The exhibited print is 5 × 3 m, at which each image is about 1 cm across. The corpus is not resolution-limited at that size: at the images' native 768 px and 300 dpi each tile measures 6.5 cm, so the same 160,000 tiles support a print of approximately 26 × 26 m.

preview/ holds a reduced image of the exhibited sort (s_highdim). The wall layout itself is not part of this release. The grid that assigns each tile to its cell is the commissioned artwork rather than research data, and it is not published. Everything the atlas was made from is here, so any arrangement can be recomputed.

Code

Not included in this release. The generation and analysis pipeline is at github.com/punktiert/DAM-Atlas.

---

Known gaps

Reported here rather than left for you to discover:

• 3 of 200,000 recordings are incomplete; those tiles are absent from the embedding tables and flagged in the tile table.
• attn_place / attn_features are all-zero arrays; that capture was deferred. The keys exist for schema stability only.
• Only one quantization was run (NF4). A full-precision comparison did not fit the available hardware, so quantization is the one untested axis.
• country labels come from the source gazetteer and are not adjudicated for disputed territories.
• The captions are one vision model's reading of another model's output. They describe the generated image, never a real place.
• The sampling plan over-allocated toward places expected to be absent from Wikipedia. In the realized sample, 10.1% of places have no geotagged article within 2 km; the intended undocumented share was larger. The distribution is reported as measured, not as designed.

What this dataset is not

It is not a record of what these places look like. Every image is generated. No photograph of any real place is included, and nothing here should be used as a depiction, a survey, or evidence about the built environment of any actual town. The corpus documents a model's imagination, which is exactly what makes it useful for auditing that model.

Stereotype in this corpus is content, not error. The model answers for thinly-documented places anyway, in confident cliché, and that behaviour is the object of study.

---

Licensing

The dataset as a whole (images, embeddings, captions, prompts, metadata and results) is released under CC BY-NC 4.0, the same licence as the earlier LCA-GCS corpus.

On the generated images. They were produced with FLUX.2 [dev]. Black Forest Labs grants users full ownership of the outputs their models generate and permits those outputs to be used for any purpose, including commercial ones; the non-commercial restriction in the FLUX [dev] Non-Commercial License applies to the model weights, not to what they produce (see the licence file). The NC term on this dataset is therefore the author's own choice for the compiled collection, not a constraint inherited from the model. If you need the images under other terms, ask.

Upstream sources. Place names, coordinates and population derive from the SimpleMaps World Cities database (CC BY 4.0); Köppen classes from Beck et al. (2023). Both are attribution-only and permit this redistribution.

No warranty. The data is provided "as is", without any guarantee of accuracy, completeness or usefulness, and no warranties are given. To the fullest extent permitted by law, the author is not liable for any claim, damage or other liability arising from the data or its use.

Citation

@misc{koehler_latent_earth,
  title  = {Latent Earth: An Atlas of Architecture in Flux.2},
  author = {Koehler, Daniel},
  year   = {2026},
  note   = {Commissioned by Deutsches Architekturmuseum, Frankfurt}
}

Commissioned by Deutsches Architekturmuseum, Frankfurt. Lead: Daniel Koehler, The University of Texas at Austin. Model: FLUX.2 [dev], Black Forest Labs.

Related

• Punktiert/LCA-GCS — the earlier, deductively-prompted corpus: 1,060,166 images of 5,856 cities across 77 architectural types and four Stable Diffusion models, browsable at largecityarchitecture.org. Where LCA-GCS asks the model for a type in a city, Latent Earth gives it only the name and lets the model propose the type.