Generative geography / Model experiment

Map 1

Represent geography as a program and explore generated map tiles in a 3D renderer.

My contribution: I developed the map-generation experiment and illustrated atlas described in the existing research material.
Tools & methods: Transformers · Vector tiles · Domain-specific language

Existing flyover recording of map tiles rendered as a suburban scene. Rendering is separate from generating the map program.

The problem

What this project investigates

Adjacent generated map tiles need to connect. A structured representation makes shared-edge constraints explicit rather than leaving continuity to the image alone.

The build

What I built

The write-up describes a transformer that produces vector-map programs one tile at a time. Existing recordings show those programs rendered as terrain and suburban scenes.

Technical choices

Decisions behind the build

Represent maps as programs

A domain-specific language describes vector features instead of storing the output only as pixels.

Make edge continuity explicit

Shared-edge contracts constrain how neighbouring tiles connect.

Current status

What the evidence shows

The portfolio contains an illustrated atlas and recorded 3D scenes. The atlas labels long rollouts, inpainting and a second building decoder as designed but unbuilt; it does not establish unlimited, reliable world generation.

Further work & evidence gaps

Longer rollout evaluation and the planned decoder need delivered evidence before being described as capabilities.