Verions 9:
I decided to leave version 8 for those who want to play with it, here is version 9.
As for samplers it should be pretty much the usual.
I mostly used DPM++ 2s a in my tests.
Version 8 experimental:
Not 100% sure yet about the samplers, should be similar to V7.5
Also careful with the negatives seems too much can make it worse.
It can be quite finicky but makes a nice film / polaroid style effect if you wan to.
2 things I noticed is it works with Euler a but if can cause issues with Adetailer.
Harder to get it to fallow background prompts it seems, you might have to force them a bit.
I guess the experimental really fits, it looks like you can get some interesting results too without the pony scoring and even some standard SDXL embedding.
Example from what happened to me, can't make elf ears, removed the pony scoring and there they where.
Version 7 and 7.5:
Euler a, DPM ++ 2M, LCM starting 8 steps (12 is better)
DPM++ starting 4 steps (6 or more is better)
DPM++ 2s a might be good too
Version 6:
Euler a Karras cfg 2 8 steps or cfg 3 12 steps
If you get artefacts lower the cfg by 0.5 or up the steps.
DPM++ SDE and LCM seem fine too.
So does DPM++ 2m SDE SGUniform
Version 2:
DPM++ SDE Karras or SGM Unifrom
Steps: 4-8
CFG: 2-3 (1 works at 8 steps, 4 seems ok after 6 steps)
Euler a Karras or SGM Uniform
Steps: 8-12
CFG: 2-3
Description
based on version 7.5
merged in 3 Loras I trained.
Goal was to change the look but not use other checkpoints or non pony Loras.
not sure however if it sill counts as semi-realistic
FAQ
Comments (14)
I like it! Getting insanely good results with no LoRA's, the results are also almost noiseless and with very sharp edges which is my style, the fact that it wors so well without detail/improvement LoRA's is also a plus for me as most tend to destroy my renders as I like to check 'em up close and notice a quality drop.
Glad you like it. I tend not to use add detail ones either, I tried to make it better with my 3 style pony Loras trained on photos. Might just have overdone it used like 2000 pictures in total.
9th is not as beautiful as 7.5, but it allows third-party lore into itself. I'll leave it, both will come in handy...
yes it all depends, even find V2 useful if I want more "cartoony". V9 is less flexible in some ways as it it slightly more realistic.
personally I am keeping v2, 7.5 and 9.
I might remove V6 too and add v7.5 vanta but I did not test that really yet.
@TiwazM More realism, man, we're all fucking realists)
@labrombula378 well maybe I will do it but as a new separate model
Is it just me, or does it seem like version 9 has a bit more trouble sticking to prompts? Example, in version 7.5 I had a prompt that includes "shirt open" and it worked like one would expect - an unbuttoned shirt. Version 9 struggles to show an opened shirt, usually outputs a full shirt with nipples poking through. Should prompting style be changed at all from 7.5 to 9?
What I noticed and I would still have to test more is that sometimes you need to "force" v9 more, like for example use (unbuttoned shirt) or even (unbuttoned shirt:1.2).
Would need to test that but I had a similar experience with colored skin for example.
seems fine for me https://civitai.com/images/12261487 or https://civitai.com/images/12261999
unbuttoned and open both seem ok.
@TiwazM Ok thanks, I'll give it a shot! LOVE the models otherwise! 7.5 was the first I tried and I was blown away! Just saw your new model too, downloading now!
Buddy, do something with the elongated heads and the very high foreheads, make the heads a little rounder. Thank you for your work.
Did you try the 2 new ones? It should be a bit better.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.