Marcelo Eduardo has been experimenting with reducing the cost of generating pixel cities by training LoRAs for a small model. After 20 rounds of automated experimentation, he shared the results and plans to tackle airports and other difficult structures next.

I thought the results were worth sharing here. Has anyone tried a similar approach to making small models work well for a specific visual style?