Question about training a character LoRA with multiple outfits
Hey, I've read your guide on making LoRAs and it really helped me a lot!
I wanted to try to make my own character LoRA with this character:
I want the LoRA to be able to distinguish between the character having wings (and the headgear or whatever) and without, so effectively I want to train two "outfits".
However, when I put the identifier/activation tag/trigger, the LoRA starts to learn the wings and the headgear and puts them everywhere, even if not specified. I have also tried separating both outfits into different folders and giving each dataset a unique tag, but the results still show bleedthrough of the wings and headgear.
What can I do to separate both styles more?
I really appreciate your help!
Hey, thanks for taking an interest. My instinct would be to treat the wings as an accessory. Make sure the images are properly tagged, and that some have the wings tag while others don't. With a few examples of each the AI should learn the difference.
Thanks, I fixed it.
I forgot to specify the caption extension, which led to the trainer ignoring my tags and mashing every concept together. BIG YIKES. After fixing that, the results were much better.