19 points juscha1 50 minutes ago 10 comments

Technical Write-up: https://www.justinchavez.com/blog/isometric-sf-how/

Hi HN! Showcasing this fun project. I took 1300 square miles of the SF Bay Area from Google Photorealistic 3D Tiles and transformed them into a cohesive pixel map. It's really fun to explore and stumble upon various famous landmarks like Apple Park, Palace of Fine Arts, or UC Berkeley.

The technicals involve using Nano Banana 2 to train a Qwen-Image-Edit 2511 LoRA which is used to prototype the style and generation methods. For final map I used Qwen to train a Klein 4B LoRA which reproduced the style at 1/4 of the cost. More details in the blog post.

Let me know what you think! I really enjoyed this process of fine-tuning smaller open source models.

If I did the map with Nano Banana 2 it would have cost around 7k USD!

Credit for the idea goes to Isometric NYC.

yboris 23 minutes ago | parent

Impressive! And the aesthetics look great!

NeuraCryption 18 minutes ago | parent

This is amazing and is 1:1 global map. Did you use google maps API to build this or UE5?

leugim 9 minutes ago | parent

Google 3D tiles.

juscha1 6 minutes ago | parent

I used the Google 3D Photorealistic Tiles API then rendered the tiles in three.js to capture them at the angle I wanted

kenferry 16 minutes ago | parent

…not actually pixel art though, no?

dnlzro 15 minutes ago | parent

I've been seeing a lot of these kinds of projects make the rounds on Twitter lately. Yours is just as good as the best of them, so great job making it more efficient without loss of quality!

But... I'm always disappointed when I zoom in and see that it's not actually pixel art, and it's full of the kinds of artifacts that are typical of any AI-generated imagery :(

scosman 14 minutes ago | parent

leugim 12 minutes ago | parent

How much it cost you paint per km2 after training the model? I did (and publish here) https://www.tinyvalencia.com and got to 0.4$ per km2 after training the model (for 15$ or so).

cmdoptesc 6 minutes ago | parent

Hi Justin, great writeup!

I really appreciate your human-written documentation because it's nice to read your thought process instead of LLM marketing soundbites.

re: orientation. I know it's creator's choice, but for me, it's either north or west towards the ocean. East makes me want to hold shift to rotate. :)

Thanks for making and sharing this!

jeffbee 1 minute ago | parent

Not bad. I feel like instead of using an image-mode model, it would be better to use a code model that generates the scene. Reason I think this is that the image generators aren't respecting the rules of "pixel art". Straight lines and flat planes are straight and flat in pixel art. These images are wavy and irregular.