Try it: sketch or screenshot → grown town
Black lines are roads, warm colours are buildings, green is kept free. Upload works with map screenshots too: dark thin features read as roads, warm blocks as buildings. Terrain is procedural here; the research pipeline uses real elevation.
Real-town mode: grow a real place on the map
Click anywhere on the map to drop a 1.9 km study window. The site fetches real elevation (AWS terrain tiles) and real roads and buildings (OpenStreetMap) for that spot, runs the model, and drapes the generated growth over the map. Five experts cover villages, hilly towns, planned fringes, informal frontiers, and megacity edges; the router picks one per window and the status line reports which.
Read the rules from your planning document
Paste text from a zoning ordinance, comprehensive plan,
or model-code excerpt (for example your jurisdiction's adopted version of
the International Building Code; no code text ships with this site). A
deterministic reader extracts the numeric development rules it
recognises, quotes the sentence each number came from, merges duplicates
under a strictest-wins rule, and can apply an extracted slope limit to
the steep lock above. This is a machine reading, not an interpretation:
confirm every value against the document. The full extractor, including
an optional local language-model layer and the compliance checks the
rules feed, is src/jurisdiction.py in the repository.
What you are looking at
The model is a mixture of five compact conditional U-Nets (~13M parameters each), one per growth regime: village, hilly town, planned flat fringe, informal peri-urban frontier, and megacity edge. A hard router on the conditioning statistics picks the expert for each window, and each expert denoises a growth plan given five conditioning rasters: elevation, slope, the existing built footprint, the road network, and water. The output has three channels: new roads, new built density, and a proposed amenity-density field. Alongside housing, it suggests where a bakery, school or clinic would let the new streets work as a 15-minute neighbourhood. In the full pipeline, sixteen futures are sampled per site and an 11-metric scorecard (walk coverage, green preservation, flood and landslide avoidance, congestion, access equity) keeps the most sustainable ones.
Example output
