As NYC wakes up to the largest blizzard in years, I’d like to share [isometric.nyc/snow](http://isometric.nyc/snow) - a wintry mix of the original isometric pixel map of NYC .

This work builds on the original isometric.nyc ideas but aims to increasingly automate the process of generating the tiled images. I’ve written up a follow-up addendum at [cannoneyed.com/projects/isometric-nyc\#snow](http://cannoneyed.com/projects/isometric-nyc#snow) that describes some of this work, but the tl;dr is:

* ensuring that training data and generations are color balanced / normalized goes a long way for consistency (duh) * augmenting water with a crosshatch pattern allowed the model to learn what it was instead of hallucinating (interesting!) * dramatically increasing inference throughput on [modal.com](http://modal.com) meant I could generate 10s of thousands of tiles in a few hours at very little cost, allowing me to experiment much more rapidly

This project continues to be a lot of fun, but I’m now mostly focusing on the agentic workflows that power this kind of ambitious generation at scale. Can’t wait to share more soon.