Now
One working slice
80 tiles generated so far, stitched into one interactive mosaic with Coit Tower and Washington Square marked.
Downtown San Francisco, redrawn as an isometric pixel map.
A fine-tuned image model redraws every block of the city as pixel art — tile by tile, then stitched into one continuous map.
The map
Drag to pan · scroll to zoom · click to inspect
The model
Fine-tuned on Qwen Image Edit
The current map was generated with Qwen-Image-Edit-2511 — a ~20B image-editing model — fine-tuned as a LoRA on about 250 render-to-pixel-art pairs (50 city views × 5 training variants) at 1024 px, trained to ~2,750 steps. Every tile is then generated with that LoRA attached.
Experiments
Cheaper generation
Today a tile takes 25 steps for full quality. Fewer steps are cheaper but quality drops. The goal is a Turbo/Lightning LoRA or Krea 2 style transfer that keeps quality at the low step count.
The map
A Coenen-style San Francisco
A city-wide map in the same spirit as Andy Coenen’s Isometric NYC — whose launch post drew 7,800+ likes. Full SF is about 14,400 tiles: 12 km × 12 km at 100 m pitch is 120 × 120.
If cost drops
The entire Bay Area
If faster few-step generation brings the cost per tile down far enough, attempt the whole Bay Area at the scale suggested by this satellite view.
Deliverables
What ships
A complete pixel-art map of San Francisco, with the training experiments, cost breakdowns, and generation architecture documented alongside it.
Why me?
I've built SF maps before
In 2024 I shipped SF Park Game — a playable map of 250+ San Francisco parks built from DataSF and OpenStreetMap data — and open-sourced the code. 10+ years in machine learning, focused on local model fine-tuning. I was the first SF hire at Inkeep (YC W23), deploying AI agents for companies like Anthropic and Postman.