Care to share the OneTrainer settings/cofing?
Thanks for the good work.
While Z-Image LoRAs were unusable for me, Klein 9B was cool, these Krea2 ones work great in my use case.
Could you share your training config for this?
He has added hundreds of celebrities, and Krea also offers hundreds without loras. Which other celebrities do you want that badly?
He has added hundreds of celebrities, and Krea also offers hundreds without loras. Which other celebrities do you want that badly?
Who said I want celebrities? And why do you care?
For your information, I create AI memes with me and my friends as subjects, and training character LoRAs for that is helpful. That's why I wanted someone good and working training config.
Ask in discord, he is active there.
He has added hundreds of celebrities, and Krea also offers hundreds without loras. Which other celebrities do you want that badly?
There are what, tens of thousands of celebrities in the last 100 years? Someone might want an Abe Lincoln LoRA, and a George McClellan LoRA to show up in the animated debate. There are vast numbers of 60's and 70's actresses and actors not generally available, even 80's and some 90's ones are missing. And I'm only talking about American and European ones, that leaves all of Asia and Australia who have their own vast celebrity history.
The world is a large place.
If somebody is looking for one, for now I recommend Automagic v3 (v1/v2 also work) and Sigmoid (or Cosine); LR 0.0001, LoKR 4 (or LoRA 64/32). For me it works great. Steps depending on amount of images, but Krea learns fast and in my testing doesn't overfit quickly.