How Pixel Berlin was made
Made by Rupert Hoffschmidt-McDonnell & Opus 5.5
Pixel Berlin is one drawing of the whole city, made by a computer from 3D map data in a muted, hand-drawn style. It was made by the AI model Claude Opus 5.5, inspired by tiny.london and isometric.nyc, and it cost about 100 dollars of computing.
Made by an AI agent
All of this was made by Claude Opus 5.5, an AI model by Anthropic, running at medium effort in Claude Code. Claude planned the pipeline, wrote every program, trained the style, rented and managed the graphics processors, watched the night run, repaired the bad tiles and built this website. Rupert Hoffschmidt-McDonnell gave the direction, made the design decisions, set the budget and paid the bills.
1. A 3D model of the city
The starting point is Google Photorealistic 3D Tiles, a 3D model of Berlin made from aerial photos. A small program on a Mac loads the model in a hidden web browser and takes a picture of the city tile by tile. The camera looks down at 45 degrees and is turned 15 degrees from north. It has no perspective, so a building has the same size everywhere on the map. Each picture covers 300 meters across, so one pixel is about 30 centimeters.
Some pictures came back blurry, because the browser sometimes keeps a rough version of the 3D model. The program measures the fine detail of every picture and takes a blurry picture again in a fresh browser page.
2. Teaching a model the style
The drawing is made by Qwen Image Edit, an open image model that changes a picture according to a short instruction. We trained a small add-on for it, called a LoRA, on pairs of pictures: a 3D picture and the same view as a gray, hand-drawn illustration. The training ran on fal.ai. A second add-on lets the model finish a drawing in 4 steps instead of about 40, which makes it about ten times faster.
3. Drawing the city without seams
If every tile were drawn alone, the tiles would not fit together at their edges. Instead, the city is split into 76 blocks of 12 by 12 tiles, and each block is drawn as one large picture. The model works on overlapping windows of 1024 pixels that move across the block, and the windows blend into each other at every step.
The blocks are drawn in four rounds, like the squares of a chessboard. A block in a later round treats the finished edges of its neighbors as fixed, so the new drawing grows out of the old one and no seam appears.
4. The computers
The drawing ran on rented NVIDIA A100 graphics processors with 80 gigabytes of memory, from the marketplace Vast.ai. Up to four of them worked at the same time, in Texas, Oklahoma and Massachusetts. A full block took about 33 minutes on one processor, and the whole city was drawn in less than a day. A small program on the Mac sent the blocks out, collected the results and kept track of which blocks were ready.
5. Color and repairs
The trained style keeps shapes and light well but removes too much color from trees and meadows. A color step gives nature back a soft green from the 3D picture, and a final grade turns the grays warm. Open water has almost no detail in the 3D pictures, so the model sometimes invented buildings or a horizon on large lakes. A second pass found these tiles automatically and drew them again with an instruction for calm, open water.
6. The website
The finished tiles are cut into a zoomable picture pyramid of about 10,000 image files, shown with the open-source viewer OpenSeadragon. The map ends at the official city border. The search knows 375,519 addresses and about 15,000 streets and places from OpenStreetMap, and it runs completely in your browser.
What it cost
- Rented graphics processors for drawing: about 58 dollars.
- Training the style and making the training pictures on fal.ai: about 40 dollars.
- Rendering, hosting and search: free.