(NOTE):

The second version is in the works— thanks to a certain hanabei for providing me with some pic for the dataset, go visit her jai page— really good taste in bots and really good taste in general artstyles!

https://janitorai.com/profiles/ecf66e01-bd19-4827-910d-30da749ddb04

💕 JAI Style 💕

(Thank you to Shuriken6637 on discord and PixAi for augmenting my dataset, he also made: Heilanshui and Mento Lora, he is very experienced— so all my thanks is for him!)

–

hi hi!

So this will be short but I will be saying stuff about the dataset, the lora weight and most importantly— how you can match it up with srefs!

First, the dataset, this is built on 10 original pictures, then augmented into 99. Shuriken had a theory about the settings being too aggressive when provided with too many original pictures— so he added 10 pictures, then augmented it, and then, boom, it worked! The style successfully converged. So if your having problems with making a lora work (other than waiting time), augment your dataset. Flip it, crop it, that's it, don't turn it into black and white XD.

Secondly, the lora weight can be 0.7 till 0.9, since I think that's the sweet spot.

And lassssstly, the sref, Storm25 said that sref and lora is very good, followed the prompt correctly and such— so thus imo, use THIS sref (you can make your own but generally use this sref.):

ZF244568

I'm done here, hope you have a good time…

(If your gonna match it up with other loras, make sure you make the weights equal though.)