v1-v3

#1
by zx96 - opened

Hello again, I am enjoying your models a lot so I thought I could ask about this one too.
I have tried v3 but it seems these are pruned, correct? which version would you recommend? are there substantial differences among them?
Thanks for sharing your work! Anime images come out crisp and lovely with V3, tried both safe and soft-nsfw.

Hello again, I am enjoying your models a lot so I thought I could ask about this one too.
I have tried v3 but it seems these are pruned, correct? which version would you recommend? are there substantial differences among them?
Thanks for sharing your work! Anime images come out crisp and lovely with V3, tried both safe and soft-nsfw.

The biggest difference between v 1, 2 and 3 is the eyes. A huge improvement from the first iteration to the 3rd.

Pruned models are the norm, they output near identical images to a non pruned while saving you multiple gb's per model.

Hope this has cleared up any questions you had! :)

Hello again, I am enjoying your models a lot so I thought I could ask about this one too.
I have tried v3 but it seems these are pruned, correct? which version would you recommend? are there substantial differences among them?
Thanks for sharing your work! Anime images come out crisp and lovely with V3, tried both safe and soft-nsfw.

Hey, just confirming Dddok's post. V1 was mostly a proof of concept, done through weighted sum merges only, and then pruned. When I got some nice feedback on my results, that's when I posted it here. V2 was redone using rebasin merging to achieve equivalent results with a bit of a cleaner merging process and an easier time getting it down to a smaller file size. Close-up eyes improved significantly at that stage, which was what got me to focus on that next. For V3, I redid part of the merge and reduced the lora component of the merge. Speaking of, if you want to take a swing at that, I'll link it below. It's pretty fun on most models, but it definitely does have some eye quality issues. Should run well on top of this model at <=0.6 weight, and most others <=0.7.

https://civitai.com/models/10155/kan-liu-style-lora

Sign up or log in to comment