LoRA trained on images from ai artist: pika baka / ピカパカ / bikabaka (same person different names)
Krea2 version has many issues in both style reproduction as well as scene consistency (it causes slight anatomy / perspective issue).
I have a few ideas on how to fix them.
But it will never be "great" since the images are ai generated by the original author on older models anyways.
Unsure if I'm going to bother tbh since no base are fun enough to spend hours fixing images.
Description
V2 was cooked in Onetrainer using Prodigy Plus Schedule Free + EMA at batch size 4 for 2500 steps at 512, 768px. I use int8 during the training, it does not seem to affect the output and speeds up the process quite a lot (this was trained in 2h)