Keep in mind that gwent art doesn't inherently have 'clean and high quality' art, like digital illustration art does.
Description
*Art does not include cart format.
Lora was created with https://github.com/cocktailpeanut/fluxgym
24 images - 10 repeats, 25 epochs = 6000 steps.
Made the mistake of using the learning rate of '8e-4', but it went fine after 4800 steps, then further increasing the quality after that.
Network dim (lora rank): 74 - Haven't exactly figured out how to insert configs from ostris into fluxgym yet, but this is the closest in terms of same-size lora outputs. Will probably need to go 80 next.
FAQ
Comments (2)
only 24 images was enough to train whole lora? looks really good! did you chose images yourselves or you found only 24? if so i can send you a link to full arts collection
I have the 3.4gb~ collection too, but the issue is, it takes way way longer, the more images are added & I'm training solely on my own gpu.
If someone were to train it on everything, it'd be better to finetune the checkpoint itself.