My first goal was to simply generate still images of characters. It took a handful of training runs to find the right setup. It had not been obvious to me that it would work at all. This was eons before techniques like LoRA made finetuning on small and arbitrary datasets easy.

→ View original post on X — @genekogan