Columbina (Genshin Impact) - LoRA

[copper gyro]

Can do Realistic, Semi-Realistic and Anime aswell.

Works best with NED (Never Ending Dream) and AOM.

[odd flower]

No way they added school shooting to genshin

[copper gyro]

[odd flower]

The name says it all

[copper gyro]

Yey

[copper gyro]

You should try it without the "eye mask" tag since usually only NED gets it right.

You should have better results without it.

[pastel wind]

ah okay

[copper gyro]

Much better

[copper gyro]

😭❤️

There wasn't much Data of her so I couldn't get it to look perfect.

Next Character LoRA will defo be better.

[strange trellis]

Historically inaccurate , but i get the joke.

[strange trellis]

Really?

[copper gyro]

Only like 350 images on Danbooru, Out of which I picked like 65, almost all the other images on other sites were bad for the dataset for couldn't be used due to the composition.

Also, Reviews Appreciated

[strange trellis]

Bad. Very bad. It’s so bad that even the creator thinks its bad. So it must suck. And it does suck. It is a whopping negative 5 stars. Do better

[copper gyro]

I don't think it's bad, think of all that Columbussy tho

[strange trellis]

[copper gyro]

Okay

I made a new version for this.

I need testers.

Trained using new method I learned. ( For Character LoRA's )

[strange trellis]

And that is?

[copper gyro]

theft

[strange trellis]

?

[copper gyro]

I just asked some people who've made good Chr LoRA's.

[strange trellis]

I mean another thing you could do is look at file information in the loras u download, theres two extensions for doing just that

[copper gyro]

yep

[strange trellis]

So u did both?

[copper gyro]

but alot of content creators don't have them

so yeah

did both

[strange trellis]

More efficient i do suppose

And it all paid off i asume?

[copper gyro]

yeah

still some things I couldn't match because not all people who train LoRA's use the same trainer

I'll do that tmrw

[strange trellis]

Ive looked at other peoples settings before, but i still havent really been able to nail a surefire method

For chars or styles

[copper gyro]

mostly because Dataset varies alot

in like the amount of images an artist or character ha

the quality of the artwork

etc

[strange trellis]

Well sure, i figured that was a given

But even then, u can still generate more pictures to use with earlier trained versions and then add those to the dataset

And train again

[copper gyro]

Never thought I'd say this

But I'm going to retrain this again

[strange trellis]

[brave pagoda]

Looks better than older version

[copper gyro]

It defo is, I'll see if I wanna upload it or not.

[copper gyro]

@strange trellis I've finished re-training this as best as I could.

Release soon.