Sunna from Zenless Zone Zero.
You might want to caption the clothing roughly to guide the model towards the base outfit or the "afternoon tea break". Explaining the details it get wrong will also help.
This lora is far from perfect, it struggles with the hairclips, and getting the correct earings. V1.1 is a lot better in that regard, but still far from perfect. I have a few ideas on how to improve it but we'll see if I am too lazy or not to implement them.
There is some light style bleeding, as this wasn't trained with a reg dataset.
Description
Trained in OneTrainer with:
Adamw, 1024px, bs 4, beta .9 .96 , orthograd, atan2 scaling, cautious variant, rank 16, alpha 1
FAQ
Comments (2)
very quick short review.
There is a tendency to create a white box on the outside of the picture with a probability of about 3-40%.
It's okay to use outside border on the negative.
A lot of images in the dataset unfortunately have white background. I might train those with mask when I do an update.