NOTE: all images in the showcase are generated using fooocus without any hires.fix or inpainting. They are all well above 1024x1024, and at weird arbitrary resolutions, as this is intended to show the way the model generalizes. Once I've trained an astigmatism lora on top this should improve further.

Just my broad spectrum custom Checkpoints I'll be starting to make going forward. I realized at a certain point for my Astigmatism line, I need a superior and consistent controllable base to train on, as otherwise I'm leaving some performance on the table.

This uses a really brute force block merge method that draws from some of my ideas regarding overfitting, and methods of combatting it. After the "pre-release" version, I will train some of my standard Astigmatism LoRas off it and roll them in, at which point we'll see a pretty standard but appreciable uplift.