This is where undercooked models will go. Not good enough to be their own thing, but interesting enough to share for anyone curious.

Depend how things will go, models from here might be deleted at any point if I feel they're not worth keeping.

First up, it's my first honestest finetune. UltronForever trained on Pins-v2 dataset.

It's a deadend project, because dataset has to be remade. For some reason Kohya SS doesn't see subfolders when finetuning, so I can't really balance the dataset like with Loras.

It knows the styles with the same trigger words as in Pins-v2 Lora. But some styles are far behind in training. 'Pencil art' has only 50 images, but repeats were the same as for 'niji art' which is over 200 images.

Still output is quite pretty, even if a little less versatile than base UltronForever.

(genitals actually degrade as I finetune it further - they still work, but upscale often ruins it)