Isometric SF is a pixel-art map of the whole Bay Area, painted as 177,000 images by a small FLUX.2 Klein model and faded together at the seams

Pan across Isometric SF and the Bay Area looks like a 16-bit strategy game that somebody forgot to ship. The site lists 25 places you can jump to, including Golden Gate Bridge, Alcatraz Island, Painted Ladies, Coit Tower, Bay Bridge and Apple Park. You can zoom, search and share a spot.

None of it was drawn by hand. The site says “Every tile was painted by an AI model from Google Maps 3D captures,” and the maker, Justin Chavez (@jus10chavez), has written up the whole pipeline on his blog. He says the job took “about 172 H100-hours from first test to final map, and I never waited on capacity.”

The interesting part is the money. Chavez says painting the whole map with the commercial image model he used to set the style would have cost about $7,000, “so I needed an open model I could train.” So he used the expensive one only for 45 style targets, then trained a 20-billion-parameter teacher, Qwen-Image-Edit-2511, to repaint darkened photos in that look. The teacher painted 1,460 places.

Those paintings then trained the student: FLUX.2 Klein, a 4-billion-parameter model about a fifth of Qwen’s size. Chavez says “Klein runs nearly 4 times faster than Qwen, which is what made painting the whole Bay affordable.” The student’s training set was 4,380 examples, which is 1,460 places times three color versions. (Teaching a small model by letting a big one do the homework.)

The color versions exist because Google’s 3D captures have color quirks, and a student model will happily copy them. Each photo went in “in true color, shifted, and shifted more, always with the same target,” so the model learned the target’s palette and not Google’s mood lighting.

Then came roughly 177,000 paintings, and a problem you can see coming from space. “Each painting is made alone, so neighbors disagree at the edges.” Chavez’s fix is a tailor’s one: each painting keeps only its center and fades into its neighbors across a thin band he calls the hem, then a color correction evens out each painting against the ones beside it. The Bay Area got hemmed, and the joins mostly vanish.

He also had to babysit an overnight job without losing his house. A small guard script polled the Modal bill and stopped the app at a set dollar amount, “so they could run overnight while I slept.” Fixing a queue problem took throughput from 29 to about 68 images a minute on three H100s at 1536 pixels, and he caught the odd mistake by eye, like a bridge drawn as water.

The tiles are © Justin Chavez under CC BY-NC 4.0: non-commercial reuse is fine with credit and a link to isometricsf.com, and the site says to ask on X for commercial use. The imagery credit goes to Google Maps Photorealistic 3D Tiles, and the map data to OpenStreetMap contributors.

The recipe doesn’t care that it was San Francisco. Any city with 3D imagery and a style someone can describe in 45 targets could get the same treatment, and the teacher-and-student trick works wherever a big model’s price per image is the obstacle.

The page defers the evals: a fuller write-up, including more on seams and color correction, is still to come. For now his own numbers leave a gap: about 119 of the 172 H100-hours went to generating the map, under 6 to training, and the page doesn’t break out the rest.

Pixel perfect, give or take a seam.