differential-studio/urban-opengen-tiles-v2

Urban OpenGen tiles v2: sixty cities from OpenStreetMap

Sixty cities on six continents, each a 10 km × 10 km window centred on the core, cut into a 25 × 25 grid. 625 tiles of 400 m per city, 37,500 in all, of which 30,957 carry built content; the rest are sea, or ground OpenStreetMap has nothing for. Every tile is 128 × 128 px at 3.125 m per pixel.

Every city was rendered by the same code with the same conventions, which is the whole point. Opportunistic crops from different sources cannot be compared to each other; these can.

• Renderer, models and documentation: https://github.com/differential-studio/urban-opengen
• Trained on this: urban-opengen-inpainter
• Companion dataset: urban-opengen-tiles-v1, the coarser set the GAN was trained on

What a tile is

A flat-colour PNG in a fixed palette, no anti-aliasing, so every pixel is exactly one class and the masks are exact:

colour	class
(0, 0, 0)	open ground, water, anything unbuilt
(v, v, v) grey	building; v encodes height, grey 100 is 4 m, grey 255 is 100 m, linear between
(255, 0, 0)	street, drawn as a filled centreline stroke on top of everything
(128, 255, 0)	greenery

What is in the download

• <City>/<City>_0.png … _624.png, the 625 tiles per city
• <City>_raster.png, the whole 10 km window as one 3200 × 3200 image, which is what the diffusion model trains on: random crops at any physical scale rather than a fixed grid
• <City>_grid.json, the georeference: centre, CRS and per-tile coordinates
• conventions.json, recording the street widths, green tags and height assumptions the renderer used

How it was made

phase0/osm_tiles.py in the repository renders this straight from Overpass, for any centre or address, so you can extend it to your own cities rather than only use ours. The conventions are deliberately simple and written next to every output: every street class 8 m wide, height from the height tag, else building:levels × 3 m, else 7 m, green from the usual landuse, leisure and natural tags, draw order green, building, street.

The same code runs at training time and at inference time, so training data and live input come from one path.

What the data actually looks like, per city

These are honest findings from rendering the sixty, and they matter if you train on this:

Heights are the weak channel, and unevenly so. Share of footprints carrying neither height nor building:levels, which therefore sit at the 7 m default:

city	untagged
Jakarta	84%
Beijing	52%
Dar es Salaam	43%
Berlin	19%

A model trained naively on this learns that Jakarta is flat. Options are to drop those cities, to mask the height loss for untagged pixels, or to accept it and say so. We accepted it and are saying so.

Feature counts do not measure coverage. Beijing returned 21k buildings and Jakarta 292k, and both are fine: Beijing's are 2,000 m² slabs, Jakarta's are 600 m² houses. Judge a city by rendered coverage and buildings per tile.

Some cities are not mapped for this purpose. Luanda returned 1,787 buildings against 8,758 street segments. The network is there, the buildings are not. It is in the dataset as rendered, and is the first city we would replace.

Coverage gaps exist. A few cities have blank bands where a fetch came back short. phase0/audit_osm.py in the repository reports per-city coverage, missing buildings and missing heights, and is the first thing to run before trusting a city.

Metrics

meta.csv carries one row per tile with grid position and every measured metric: coverage, floor area ratio, mean and max height, building count and mean footprint, street length from a Zhang-Suen skeleton, intersections, dead ends, block count and mean block area. The same functions measure generated tiles, which is what makes model output comparable to real cities by numbers rather than by eye.

Uses

Trained two generative models. Also usable for morphological comparison between cities, as a segmentation or inpainting benchmark, and as input to the sun, view and density analysis in the repository, which computes direct sun on the ground, street and walls at each tile's own latitude.

Licence and attribution

ODbL 1.0, © OpenStreetMap contributors. This is a derivative database of OpenStreetMap. If you use it, credit OpenStreetMap and, if you would, us:

Contains information from OpenStreetMap, available under the Open Database License.
Urban OpenGen tiles by Differential (differential.studio).

Known issues

• The height default of 7 m for untagged buildings, quantified above.
• A flat 8 m width for every street class, which is a simplification, not a measurement.
• A handful of cities with fetch gaps; see the audit.
• Nothing here has been checked against ground truth. It is OpenStreetMap rendered consistently, with all of OpenStreetMap's own unevenness intact.