One Epoch trained on 29 images, first attempt, 768x768 training resolution, 2900 steps
GC-v1-anime is trained on 36 images, 100th epoch, same resolution, 10 repeats per image, roughly 5000 steps.
Recommended usage is at 0.75 strength, tendency to have extra ears/arms/legs so might need to negative them out.
Description
v1 was never released because it was pretty trash, will improve, issues include getting multiple ears on bunnies which might have to be img2img'd out, recommended usage to use (multiple sets of ears) in negative prompt
FAQ
Comments (14)
What exactly does this change with images?
incremental added to images
@PsySpy i thought it would help improve anthro ponies in my lucidpony model but not seeing any changes at all, no matter the strength. is it only for the yiffy model?
@cocoyoc it shouldn't be, I've just noticed Yiffy has the best results that I've seen. It was built as a base 1.5SD LORA and works on other models I've tried though how much it works might depend on the model. Downloading LucidPony to test it.
@cocoyoc Just tested it with LucidPony 2.0 and it does work, pretty obvious changes, uploaded a partial incremental to images. Not the best quality coming out but that could also be my prompting.
@PsySpy maybe i just need to keep experimenting
apparently my Beta Ui's command window is giving Lora errors, so it may be a problem with the UI, not the Lora, itself. Will get back to you after some figuring out
@cocoyoc It is a Kohya created LORA if that makes any difference, I don't think it does though as I think most people use Kohya to create LORA.
@PsySpy No, it's all LORAs giving the error, not just this one. Definitely a UI issue
@PsySpy Got it working :-)
When it got good then let's me know I definitely need a hand in production XD.
So far what it really captured my style is how I do background but the characters still very off. So.. Good luck with the effort.
I can't believe you replied to this xD
I agree that the lora is not very good, moves towards the style but didn't capture it. I didn't use nearly enough examples and of too many characters. If you don't want it here I'm perfectly fine removing it since you're the artist it's based on.
@PsySpy Nah no problem I personally have nothing against this. But I kinda still wish people asking the artist first before they using their works for training the AIs.