LoRA Trained on SFMTA CAD Drawings to Aerial Images

Hey I'm Kieran and I've been playing with the intersection (pun intended) of generative AI and civil engineering for roadways. I did a test run training a LoRA on the new Flux 2 Dev model using Fal's trainer useing a custom dataset of paired images from publicly available striping CAD drawings of street layouts to aerial images of the same area.The use case here is to allow urban planners to instantly visualize their proposed changes as they work with their existing tooling.This was just a quick experiment with a small data size that exceeded my expectations so I wanted to share with you all.Watch a demo with instructions on how to test it out: https://www.youtube.com/watch?v=zS8pGoOfe00Try it now (includes free credits) https://3dstreet.app/generatorIf you're really excited about running on your own hardware here are the lora weights: https://v3b.fal.media/files/b/0a87f41f/glySGbKtv8lzigPWzQDjb...I can writeup a longer blog post if interested in the details. This was only trained on 12 image pairs with text descriptions but it still cost about $100 on Fal. I'd love to do a larger run, but it does take a while to prepare all of the data and I'm hesitant to drop $2k. I'd be curious for the experts out there if you think the quality will increase if I use a larger sample size.