>>27787 2M + SGM Uniform works quite well for me with APG. AYS too, but it's a little muddier.

>>27786 i always assumed the retreat to euler was due to shitbakes that couldn't resolve a higher level of detail properly dpm seems to frequently kill color with kohya hrfix though

>>27789 Not too sure about that, I never used NAI much but I see it quite frequently in the metadata. It isn't just for SMEA, either. I don't notice loss of color info using regular hires fix (yet). Also just noticed vpred 0.6 can finally get some decent results with this object insertion, just the tip prompt.

Any tips on giving the dude in hetero couple pics long hair out of the box (before inpainting)? Especially when both share the same space or overlap, like the girl being in front of the dude etc.? I tried regional prompter, but that doesn't really work even with long hair having large emphasis, same for short hair in negatives. It's like the short hair is inherent to couple scenes and can't be overwritten.

https://mega.nz/folder/ctl1FYoK#BYlpywutnH4psbEcxsbLLg added alkemanubis (noobvpred05) added atenaba (noobvpred05) added fuyubone (noobvpred05) added ootsuka shinichirou (re zero artist) (noobvpred05)

>>27786 based

>>27793 How could this be "by (alphonse \(white datura\):1.2)"? I assume it's the pokemon art style leaking? The aspect ratio in metadata also doesn't match, did you crop it? >ControlNet 0: "Module: inpaint_global_harmonious, Model: Kataragi_inpaintXL-fp16 [d20c87b4] Is it better than ordinary img2img inpaint, based on your experience?

>>27795 It really is, pokemon is leaking hard

Crossposting from 4ch/h/. Some saturation comparisons on Noob vpred 0.6 (itercomp perp merge, but it should behave mostly the same as base). 1. 6 APG: https://files.catbox.moe/vkt851.png 2. 4 CFG, 0.7 rescale: https://files.catbox.moe/8699i5.png 3. 1.5 APG, CFG++ sampler: https://files.catbox.moe/7vmv46.png I like the last one best. Snakeoilmaxxing.

>>27797 what's wrong with the dude jfc

>>27798 It's Muk, the Pokémon. Actually makes an interesting test based on config of how strong the negatives need to be to stop it from leaking.

>>27799 >It's Muk, the Pokémon oh oh that's fucking hilarious

>>27800 Doing "artisr:muk \(monsieur\)" actually helps a lot like it does on NAI, but I forgot to do it for those tests ;_;

>>27797 apart from dude turning into a slime, I like cfg+rescale (2nd) the most visually

>>27797 I believe that CFG 1.5 with CFG++ samplers is equivalent to CFG 10 with normal samplers The team behind CFG++ used 0.6 in their examples, anything above 1.25 gives me hard outlines with v-pred

>>27803 It's 0.5 CFG before scaling because 1 CFG in webui is actually 0, iirc. It doesn't let you set it lower.

>>27804 Didn't about that, thanks

>>27797 >Snakeoilmaxxing. why do people keep calling cfg++ snake oil? there were unironic improvements with it when using illustrious 0.1

>>27804 Wait, it's a setting you can adjust in ui-config.json

>>27807 I'm only joking because stuff like SAG and freeU basically gave no improvements in the past despite touting themselves as straight upgrades. Both APG and CFG++ samplers have had plenty of cases where they are demonstrably better.

Do you guys use hyper/lightning/turbo or whatever these are called? I'm testing them with illust-noob shitmixes and wonder if I'm doing it correctly.

>>27810 ...what?

>>27795 >I assume it's the pokemon art style leaking? yep, lana in particular pulls the official art style hard due to her simple face >did you crop it? that's from inpainting her face, genned it at 1536x1920 kohya hrfix 1.5 downscale 0.39 step >Is it better than ordinary img2img inpaint, based on your experience? much, it's a lot more context-aware. i'd say it's about 60% as good as NAI's inpainting and that's a high compliment.

she deserved it

>>27812 Thanks, it's night and day in the few tests I did

I love being able to lewd OCs with this technology

>>27816 Oh good it's not just me

>>27818 have you somehow managed to fix karras on vpred?

>>27819 nta, karras converges at 60 steps

>>27820 >60 steps jfc, I wonder if that's pure sampler weirdness or if vpred really is severely undercooked. I was doing 28-30 with euler a + SGM uniform and things come out fine, switched to dpm++ 2m and was able to push it a bit lower. admittedly I do 12-16 steps of hires fix on anything I actually post which is extra denoising, but even on a 3090 anything past 30 steps is way too much wait time for base gens

it's bugging me so i did some single-seed tests 1. dpm++ 2M karras 2. dpm++ 2M SGM Uniform the latter reaches a "clean and coherent" image much faster

>>27823 >(childish pussy:1.5) >stuffy smell >cumpussy, cumanus >kinomoto sakura wearing white thighhighs garter belt with orange (brown middle short hair:1.6) intakes and green eyes, (daidouji tomoyo) wearing black thighhighs garter belt with blunt bangs purple (black:2.2) (long hair:1.6) and purple eyes and fair skin,kinomoto sakura wearing white thighhighs garter belt with orange (brown middle short hair:1.6) intakes and green eyes lifting one leg to show the soles👣