Thread 8689703

Erufu breeding Edition

Previous Thread: >>>>8681447

>LOCAL UI
reForge: https://github.com/Panchovix/stable-diffusion-webui-reForge
Comfy: https://github.com/comfyanonymous/ComfyUI

>RESOURCES
Lora Training: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainer
ControlNet: https://rentry.org/dummycontrolnet
Inpaint: https://rentry.org/fluffscaler-inpaint | https://files.catbox.moe/fbzsxb.jpg
4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb

>FAQ
>What model is everyone using nowdays?
Any of those
- NoobAI-XL (NAI-XL) (base model) | https://civitai.com/models/833294/noobai-xl-nai-xl
- 102d custom (noob base mix) | https://civitai.com/models/1201815?modelVersionId=1491533
- 291h | https://civitai.com/models/1301670/291h
- r3mix | https://civitai.com/models/1347947/r3mix-illu01nai-v-pred

>How do I do X
Given how the models work and are trained, you should always look for a corresponding tag in these places
- For anime -
https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag
See tags in action: https://tagexplorer.github.io/#/
- For furry-
https://e621.net/tags | https://e621.net/pools

>What I want to do doesn't have any tag or tag related to that
Time to train a lora (see the resources section)

>How do I add/remove/fix things
Inpaint (see the resources)

>I want to do X interaction and/or put Y amount of characters at the same time
You can try to prompt accordingly for it but you will end up using controlnet and inpaint (see the resources again)

OP Template/Logo: https://rentry.org/hgg-op/edit | https://rentry.org/hgg-opV2/edit (new and open to suggestions) | https://files.catbox.moe/om5a99.png

Only 25 a month for https://novelai.net/

>>8689703 (OP)
Box for pic? And don't say it's deadnoodles again.

it's over we got colonized by /g/ vermin
i'm outtie

why do we hate nai

i love nai

>>8689706
>And don't say it's deadnoodles again.
It's not (this time)
>https://files.catbox.moe/hmkx3d.png
But to be fair it was a very very lazy almost i2i from this pic
>https://danbooru.donmai.us/posts/3807472?q=parent%3A3807472

>>8689716
Oh you know that is tiangliang. Surprised you even got tits that big when he's at 0.75.

>shitmixes in the OP
yeah nah fuck that shit

>shitmixes bad narrative AGAIN
Get new material

>something that was true
>something that remains true
>something that will always be true
>nooooo I don't want to hear it!
keep burying your head in the sand, retard.
Either way this is supposed to be a general aimed more towards training - and anyone training on top of a shitmix should find their nearest bridge and jump directly off of it.

>>8689721
Why would telling people to gen on 102d mean they should train on it?

>>8689725
>What model is everyone using nowdays?
doesn't say shit about what they're used for. Follows to reason that a thread directed towards training would only list models people are using to TRAIN on in the OP.

>>8689727
Kek, yeah ok, noted for the next thread bake

Do I REALLY have to use comfyui to use Wan 2.2

d-diffusers?

>>8689727
but this isn't a training thread? it's just a thread that happens to have a lot of trainers

>>8689784
It's a thread for avoiding one spammer who's been gone for months. People just didn't want to let it go so it kinda got repurposed into lora baking general.

>>8689809
I agree with >>8689784 we just happen to have a higher rate of bakers than genners
>inb4 post gens then

>>8689813
>post gens
Baking unironically but will post when done.

i don't post gens because my gens are boring

>>8689830
This too unironically.

>>8689830
I have a week off and I was supposed to spend it learning controlnet and drawing doodles to do some fancy composition but I've spent it gooning to pov as usual.

>>8689830
Porn is boring, erotism is the real deal

>>8689850
>ass fangs
ero...

16 channel vae... please...

>>8689830
Pretty much. I'm baking pretty often, but all I'm really doing afterwards is running the loras through testing using the same old testing prompts I have had for a while.

>>8689883
catbox please?

>>8689898
https://files.catbox.moe/9dezwm.png

>>8689899
thank you. very nice

>>8689899
where is that model from?

>>8689900
no problem
>>8689901
its my custom merged one
i don't really plan on releasing it because it's nothing special
you can probably just use base noob vpred and you'll get pretty close

>>8689883
nice, I like it

>>8689899
You Comfy spaghetti monsters still impress me

i'm tempted to bake another one because the op of this thread had a brain tumor

>>8689964
How so

>>8689965
>59xed3
>shitmixes
>furshit
>furshit
>retarded faq with questionable advice
>controlnet collection removed
>upscaler db removed
>https://rentry.org/Copier_Method_Notes is not added

actually we should rename this thread to /htg/ hentai training general

we should rename this general to /hrt/

Where's catfag?

hopefully dead

>>8689979
Burning in Hell, God willing.

>>8689975
>59xed3
no idea what is this
>shitmixes
all aproved and good for genning
>furshit
the inpaint guide should be replaced with something more up to date and with anime examples, yes, but the furry booru has very convenient tags for composition so it's worth keep it around
>retarded faq with questionable advice
give me better questions and/or better answers
>controlnet collection removed
you only need 1 controlnet model now (xnir)
>upscaler db removed
correct, should be inside or complete replaced with a proper upscale guide rentry
>https://rentry.org/Copier_Method_Notes is not added
I legit have not seen this one before, I'll add next time (if do end baking again)

>>8689902
i tried to get similar but the style looks way different

>>8689985
>you only need 1 controlnet model now (xnir)
xinsir is the person, it's union and/or promax union. I would still keep anytest as you can't get the same mix of line/normal/depth without stacking three CNs on top, would take crazy amounts of vram. Works much better for pose transfer without picking up colors and style.

And even though union does everything, you still need to understand which controlnet you want it to act as and then use the correct preprocessor. So the guide still has value, but should perhaps be updated to mention this specific CN can do everything. There's also the question of how good it is at each task compared to specialized ones.

>>8689996
>There's also the question of how good it is at each task compared to specialized ones.
Of course but, are (You) going to do those comparisons? is (He) going to do them?
I am not saying promax union is the best out there but it has worked for me for about anything I want to do so, yeah
We can add anytest there too if you feel it's worth it

i do not feel so good naisisters...

>>8689985
>all aproved and good for genning
The only people who use shitmixes are braindead retarded.
The only people who actually recommend using shitmixes are so braindead retarded that they don't even remember how they learned to actually do anything the correct way with SD shit to begin with(if they even actually did).
Not only are shitmixes bad due to being overfit biased garbage, they're bad because it teaches new users bad, incorrect and, frankly, extremely dumb fucking behaviors.

>>8690026
>Not only are shitmixes bad due to being overfit biased garbage, they're bad because it teaches new users bad, incorrect and, frankly, extremely dumb fucking behaviors.
you are too harsh. shitmixes are good as training wheels so one doesn't get overwhelmed while starting out. it's just that with experience I believe one should strive for less overfit models and ideally go for the base model, with noob it's certainly feasible.

>>8690027
the biggest part of starting out for me was exploring style mixing and getting things to look how I wanted. Using a shitmix pretty much completely destroys that experience and moves people into a mindset of 6.9GB loras instead of, you know, actual loras - which the perception of gets damaged because, lol, loras will often not have as much style transfer or even just make things look worse if you're using a shitmix.

>>8690029
>the biggest part of starting out for me was exploring style mixing and getting things to look how I wanted
plenty of people start out completely clueless and they just need the model to gen at least something coherent. I've seen newbies complain that their outputs are completely fucked with noob, but as soon as they switch to WAI they get satisfying results.
I think it's fine to use for the first week or two, it's not like it's completely incapable of styling. It's just that the mindset should be to switch to something less overfit, but most people stay on those slop models forever. I guess laziness is expected when genning coomslop

How does one create a shitmix actually? Like, what if I want to mix 15 of my loras into one model?

i tried chroma again and the hands are still fucked

>>8689985
>I'll add next time
a good copier method guide should be around 100% about how to train the base style lora because that's by far the most important part and i think that every guide I've seen talking about how to do it is wrong

Anyone know how to colorize manga scans? I tried Lineart CN but the results suck

>>8689703 (OP)
go ahead https://rentry.org/yahbgtr
i'll be spending about 1 hour every day on it, for now it's enough

>>8690120
It's the same people in both threads

>>8690120
Flux Kontext

>>8690121
Very nice, noted for the next bake

whats the difference between this thread and /hdg/?

I don't expect any sympathy or help but I'm pretty new to this stuff. Have been playing around with noobai on both the civitai webpage and locally with model+lora and the... engine or whatever just shits itself in confusion if I try to prompt 2 or more characters. I understand there's something like regional prompting but just setting this shit up so I could prompt stuff with my GPU was already a massive undertaking I have no idea how to insert this new concept into a premade workflow. I've love to generate things like this >>8689883 or just people having sex in general.

>>8690163
The easiest way you have to let anyone help you is to upload your image here:
https://catbox.moe/
So we can look the metadata of it and see what you could possibly be doing wrong

Why the fuck if you apply a lora, it's also applied for negative prompt?

>>8690166
https://files.catbox.moe/kz5g6q.jpeg
Here, I generated a few images with the same settings and this was the best I could get but still not what I wanted. Errors: the penis does not go across the face and instead goes up / lays penis AND balls on girl's face ie. the girl's lower face is not visible at all, more than one dude, a dude's bald head covering the girl's crotch, panties covering the girl's eyes while some ghost dude jerked himself in the background?

>>8690185
https://files.catbox.moe/pofkn0.jpeg
here's this other pose I wanted to generate but no matter how many times I redid this one the condom is ALWAYS the wrong way around. image related is what I wanted to achieve

Cool tool released recently, might be useful for current and would be trainers out there.

• slap on autotags for you in one click
• pull up Danbooru wiki so you know what a tag means
• drop a solid background behind those pesky transparent PNGs
• your other usual stuff inside a generic dataset editor

https://github.com/Jelosus2/DatasetEditor

>>8690167
It's applied to the Unet and the TE. You could use that modified TE to encode only the positive prompt and use your original TE for negatives, it's trivial on Comfy. Obviously not relevant to loras that didn't train the TE at all, which is the majority afaik.

>>8690197
Can it populate mask images for masked loss training? Always a pain having to create them manually for images where I do not intend to mask anything.

>>8690185
You need to be more specific with your prompts, you are also using "_" (underscore) for no reason at all
Is this want you want to do?
>https://files.catbox.moe/5pz7et.png
>https://files.catbox.moe/s1k4n1.png

>>8690163
>I've love to generate things like this >>8689883
But you are linking an image where only 1 character has detailed description, that's done without regional prompter, just describe the girl and the sex act, the male is a generic "head out of frame", or you can use solo focus, or faceless male, or whatever. Most tags just automatically go to the girl. If that's a problem - might be regional prompter time (or you may come up with something clever).
If you have 2 or more girls (or boys), current models will (usually) assign them correct eye color, hair color and clothes with just one tag (their name), so you don't need to specify these tags. But if you want some of them wearing clothes that aren't their official clothes (tm), or have them in different states of undress, specific to each individual person, or have each person do a separate specific thing, that's when you need regional prompter again.

>>8690202
Doesn't support masking. Asked the creator and he considers masking to be niche.

>>8690205
Yes, that's exactly what I wanted. I don't know how to check metadata or whatever so I'm just gonna have to try harder.
>>8690207
Thank you for your input as well.

>>8690215
>I don't know how to check metadata or whatever so I'm just gonna have to try harder.
I don't know about the civitai generator but assuming you are forge/reforge for local you drop the image on the png tab

>>8690218
I'm using comfyui actually

>>8690220
https://sprites.neocities.org/metadata/viewer

>>8690202
https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbw
i use the mask_full.py script from this folder for that. both of these scripts use the alpha channel to generate masks so if you feed mask_full a bunch of opaque images then it'll shit out a bunch of white ones with matching resolutions pretty quick

>>8690221
Bookmarked, thanks. I will fuck off now.

>>8690222
oh cool, thanks!

>>8689996
So uh... where were you when we were discussing all this stuff last thread? Are you that schizo which was attacking me instead of providing useful suggestions since this is literally what I said should be in the OP? Will anyone write a guide (I don't mean a rentry just a basic explanation like you provided here) of which model to use to do what in CN? That will go a long way to improving the OP. But of course it's a lot more fun to shitpost here and let it fade into the archives than to do something that lasts.

>>8690186
Box

>>8690026
So what are the good behaviors?

>>8690045
I forgot how to do it with loras since I used to do it in the bmaltis gui back in the day but there's also a checkpoint merge which is a different thing. You can do that in the corresponding tab in reforge.

I've been using the new anytest cns for illu. They work quite well and aren't that heavy. No clue what's different between version A and B tho

>>8690215
A good rule of thumb would be "prompt what you can see." In this case a simple "from above" would have gone a long way. Of course you're also prompting on base noob as a newbie but our resident schizo will never tell us how to use it, only shit on people for recommending you switch to either of the two models in the OP to have an easier time.

>>8690222
Can this be added to the OP under training?

>>8690120
If you want 1:1 a whole page it's not going to work, you can use it as reference so you can upscale and then inpaint everything to make it look good but the final result will always be slightly different from the original

>>8690233
>Are you that schizo which was attacking me
I don't think it's me? Not a fan of personal attacks in general, and I can't find it.

>check last thread
>crtl+f (you)
>120 results
Shit, there really are just like five of us here.

>>8690197
Thanks will give this a try on my next bake. New gacha waifu is coming out soon anyway.

someone test using the copier method with the style lora trained via different kinds of lycoris networks for me thanks in advance

>>8690200
>Obviously not relevant to loras that didn't train the TE at all
That's what I was asking about, actually. For some fucking reason every single UI applies the lora for both conditioned and unconditioned unet calls. If the whole point of the CFG is to get rid of uncertainty, WHY THE FUCK would you add certainty to uncond by applying a lora to it?????? Fucking fuck, this is nuts. This actually can make your gens with loras noticeably better. We may have been frying shit out of our loras all this time.

AAAAAAAAAAA SHIT WHY DID NO ONE NOTICE THAT BEFORE?????? IF YOU IMPLEMENT THIS PROPERLY YOU WILL ALSO BE ABLE TO APPLY ANY LORAS TO NEGATIVE PROMPT AS WELL

>>8689962
funnily i try to make my comfyui spaghetti as simple and streamlined as possible because i see most workflows where people try to do everything at once and i get a headache
>>8689987
this style isn't really all that consistent on my model either so that's probably just how it is
>>8690220
If you're using comfy then you can just download this >>8689899 and drag it into comfyui
the metadata is all in there

>>8690329
dude has an extra finger

>>8690332
all the better for fingering vocaloids

>>8690333
Yeah, he'll need them for that double cervix vagina

Please excuse my retardation. I'm attempting to run a local using comfyUI since allegedly it's the most beginner friendly. I have my "checkpoint", I have a Lora for a specific artist and a VAE (because the premade workflow I'm using doesn't allow me to delete it). I'm following the checkpoint instructions and using the recommended CFG, etc. but the end result always looks like a randomly color splooge on a white canvas. Doesn't look like anything. Why is this happening?

>>8690296
>>8690302
IT LE HECKING WERKS AOAOAOA

>>8690356
>I'm attempting to run a local using comfyUI since allegedly it's the most beginner friendly
It's not kek
>the end result always looks like a randomly color splooge on a white canvas. Doesn't look like anything. Why is this happening?
If you are using a vpred model it's probably missing the advance sampling options that make it work
I would throw away comfy all together and use reforge instead even if it's """dead"""

God, I loved abusing this girl. I'm such a dick.

https://files.catbox.moe/3wdqzr.png
https://files.catbox.moe/8o55y0.png
https://files.catbox.moe/ksvf5g.png
https://files.catbox.moe/858l68.png
https://files.catbox.moe/pyegsn.png
https://files.catbox.moe/dqws50.png

>>8690357
What exactly are you doing? I'm not really getting any big differences using loras only on the positive

>>8690357
>>8690367
Turns Comfy has a hook nodes, which just happen to let you load a Lora only for positive condition pass, testing it on this lora >>8684031

Using some hot danbooru prompts

>>8690360
hmm... I'll try to work backwards (from the lora to the checkpoint) and check for compatibility issues. I think I'm not using the right one and I would really like to get rid of that VAE node. The UI just gives me an error when I do though. Thanks.

>>8690370
How the fuck do you get actual spanked ads like that. Every time I try it looks weird or has handprints I just want a nice red set of cheeks

What convey newcuties to use (un)comfyUI as first UI option for local?

>>8690369
wait if i'm understanding this correctly can you also mask a lora so you can only enable it on part of the image?

>>8690369
When I said it's trivial I meant doing it like this. Is that different? It looks like you're not loading the Unet part at all.

>>8690373
everywhere else treats cumrag as a braindead easy option since they don't actually do anything other than copy/paste someone else's workflow and never touch anything ever. Also everywhere else is still obsessed with garbage like FLUX or falling for the fake hype around chroma and you have to use cumrag for that.

>>8690373
Sounds more like a last resort in this case, can't run the experiment anywhere else.

>>8690373
Reforge you have to do a bunch of other additional shit in order to install. Comfy it's just unpacking the thing into a folder and go.

>>8690385
reforge also had a just unpack option last time i checked
or was that normal forge?

>>8690385
>Reforge you have to do a bunch of other additional shit in order to install
Reforge has quite literally 1 click install

>>8690388
>>8690389
FUCK

>newcutties can't read
marvelous funny times ahead fellow /hgg/ers

>>8690370
which one is supposed to be the good one?

>>8690370
This is only relevant for loras that trained TE, right?

Yes, using loras for negative also appears to work...
>>8690372
I'm not doing anything special aside from using that lora only for positive cond, just copied tags from here https://danbooru.donmai.us/posts/9788005?q=order%3Arank
>>8690378
I think this is not an image mask, it's likely a prop mask, but it's green so idk
>>8690380
I don't load either clip or Unet directly because it'll also load both for uncond/negative prompt, Comfy loads these hooks at runtime it seems
>>8690394
Neither since I'm just testing

I wish you would all go back to /hdg/, this thread is my and noodleanons love nest

>>8690398
No, it loads both at runtime.

>>8690400
this is the lora baking general now
anyway /hdg/ is actually pretty comfy today

>>8690363
Nice. Sluts like these are great for alley fucking.

>>8690399
And if you just use ConditioningZeroOut on your negs, as we're been using on noob anyway?

>>8690373
The actual correct answer is that since /h/ is no longer newfag friendly, they've returned to /g/ or reddit or wherever else. Those places heavily shill comfy for everything and all the newfag guides are written with comfy examples/workflows etc. If they used reforge they would be significantly more confused or ignored when asking for help so they just stick to what works. It's like microsoft and apple at this point.

>>8690394
Man I should really start going to /e/...

>>8690378
nta but yes, it's a hacky way to restrict lora to specific regions. POC: https://litter.catbox.moe/2t89sf1e96ff61dj.png
That's mantis-X added on the right side

It's incredibly slow for some reason, I'm getting 50secs per gen instead of 5

>>8690422
good luck out there

>>8690424
b-box for both...?

>>8690423
good to know
i'll try to remember this in case i ever need it

>>8690412
Why, this is actually an interesting question. This still works because CFG works in latent space. If you set cfg to something other than 1 you need 2 calls to unet. And since USUALLY you apply the same lora to BOTH unet calls, you get lora in both negative and positive prompt >>8690167 which is something you actually DON'T want for CFG to make magic >>8690296
If you apply hooks this way >>8690369 the lora is applied ONLY for cond, and base model without lora is used for uncond.

>>8690425
https://litter.catbox.moe/y7jvvn160fyh51fk.png
https://litter.catbox.moe/y7jvvn160fyh51fk.png

>>8690412
So yeah, if you zero out the condition, you will only zero out the clip embeddings passed to the model. It's up to the model to decide how to interpret zeroed out embeddings. And guess what, unet with a lora and without a lora interpret zeroed out embeddings differently.

>>8690433
https://litter.catbox.moe/lgz2smtkxdgwwpkh.png

>>8690416
It's so over, why people dislike simple yet effective things again?

Is the catbox extension acting weird for anyone else? Pictures are failing to load for some reason.

what does ConditioningZeroOut even do?
it just looked weird last time I tried

whats the difference between catboxxing an image and just posting it here?

>>8690434
And your assumption is that this different interpretation is strictly worse every time

>>8690447
posting an image directly removes the metadata
catbox keeps it

>>8690447
some images are against the rules so they get deleted or get you banned so, also this >>8690449

>>8690445
You put conditioning tensor in it, it becomes zeroes.
[0.2, 0.3, -0.5, 0.9] -> [0, 0, 0, 0]
>>8690448
CFG *requires* for the uncond to be uncertain. Like it fucking doesn't work otherwise. And if you apply a lora to uncond (which is what everyone is currently doing), you are reducing the uncertainty of the uncond, by a fucking lot. It isn't a matter of being worse or better, it's just how CFG works.
This also nicely aligns with this autoguidance paper https://arxiv.org/abs/2406.02507 if you think of "unet with lora applied" as "fully trained model" and of "unet without lora applied" as "a smaller, less-trained version of the model"

>>8690454
If you're looking for an improvement it has to be a matter of being worse or better. I mean it clearly still produces images, even if you say it doesn't work.

yeah just fries the image for me
good to know

>>8690329
as a A111 user, all comfyui users are Italian

>>8690466
i wish i was italian my pasta always comes out like shit

>>8690459
These are the same people calling you stupid and laughing at you for believing your empiric tests rather than their theories btw.

>>8690469
It's because you don't fry the tomatoes for long enough. Their flavor only arrives when it fully cooks in the oil. It goes from tart to sweet.

>>8690459
Well you blurted out something vague so I decided to explain more. You know, being nice.
>I mean it clearly still produces images, even if you say it doesn't work.
That doesn't really say anything, you can produce images with CFG even if you don't do anything special. It's just that you don't see the full effect of it.
>If you're looking for an improvement it has to be a matter of being worse or better.
You can look for yourself which one is closer to the actual artist https://danbooru.donmai.us/posts?tags=rladnr6
Removing the lora unet changes from the negative and putting it only in positive clearly works, and this doesn't really have to do anything with zeroing out the condition. Plus Comfy patches the model on the fly so you don't even need extra vram compared to regular lora loading or anything like that.
>>8690473
Please keep your headcanon to yourself.

>>8690479
He's not entirely wrong. I don't know what cfg does under the hood, beyond what I've managed to intuit while genning anime porn. Never needed to know and likely never will. I assume that goes for most people here.

So it's easier to understand we're talking about improving lora accuracy, rather than cfg not working. When it clearly "works" the way it always did.

>>8690484
The way it always worked is that it relied on lora gods, not on CFG.

I'm dumb. How do you "reset" a workflow in ComfyUI so you can run the same prompt again?

i really wish that hook lora thing didn't halve my gen time
hopefully it'll be faster when it's out of beta
>>8690511
i don't know if you can other then changing something

Damn, it's even stronger at 0.6 weight than a normal lora at 1.0. Seems like you really can just cut your training time almost in half.
>>8690516
>i really wish that hook lora thing didn't halve my gen time
It doesn't halve it for me?

>>8690521
it's slow as hell for me
my gens already aren't quick since im on 8gb but it goes from 44 seconds per gen to 90 when i enable it

>>8690526
her age?

>>8690528
Your age?

>>8690529
52

>>8690528
Old enough officer :^)

>>8690530
10

>>8690522
Just looked at vram usage, it spiked at 17 gb. Seems like comfy caches all three models for me, weird. It shouldn't really take any longer than usual or require more vram though. Lora is reapplied each step during training, so this operation shouldn't take long at all.

>>8690534
i will chalk it up to comfy bullshit and not dwell on it

>>8690538
Well? We're waiting,

>>8690542
its nai

>>8690548
it's kaamin_(mariarose753)

>>8690540
patience. She's debating.

>>8690549
and nai

>>8690526
Erotic.

>>8690433
could you reup these please?

why do you keep embarrassing yourself?

>>8690573
nyo

Are video gens any good for 2D? I see it's getting momentum on the 3DPD side of things.

>>8690626
Wan 2.2 works very well on 2D.
Look up the うごイラ tag on pixiv.

>>8690626
Yeah it do be like that. Barely okay for 2d (with lots of work), while outstanding for 3dpd.

>>8690636
nah they still look like weird rotoscoped garbage

>>8690636
Yeah I was actually browsing that tag just now.
Some stuff looks pretty ok:
https://www.pixiv.net/artworks/133675728
https://www.pixiv.net/artworks/133848568
but still far from what I'd ideally want though.

>>8690645
I mean, it's the best thing out there for now.

>>8690399
Yeah I do too and this is what I get most of the time or a red handprint.

How does he get genitalia this detailed?

>>8690713
clitoral hood, vulva, spread pussy, (erect clitoris:1.3)
-gape

>>8690714
Say if I were to try to copy his style completely, what kind of tools do you think he uses?

>>8690717
just download any version of wai and prompt nyantcha + shiny skin youll get 80% of the way there

euler a, some post processing, model looks like illustrious idk, whats the character name?

>>8690721
What is this?

>>8690721
its the doll from bloodborne

ill play around with it and get back to you

can somebody tl;dr this lora conditioning yap and tell me how i should gen with noob on comfy?

>>8690743
Lora is being put in both positive and negative prompt (without your consent) and you're getting btfo'd by the result. (You) need to raise your IQ a few points so that you can figure out how to make it apply only to positive as intended.

>>8690744
well whatever, fuck loras anyway

>>8690721
wtf is this schizo analysis

https://civitai.com/models/131405/plain-doll-bloodborne-lora
https://rentry.org/937c9cn4

>>8690792
>1.5 lora
>pony prompt
>civitai tier quality tags
What are we doing here

which reforge branch should I be using? Panchovix keeps adding more and it makes me want to slap him.

>>8690804
That's my lil nigga you speaking of.

>>8690823
sometimes the lil nigga needs a slap. I respect him for being the only light in the darkness of ComfyUI, but holy shit his branch management is ass. Seriously newforge_dendev or newmain_newforge though? They seem to be the most up to date but they seem to have different capabilities.

>reforge now refuses to inpaint images I paste into the canvas throwing some PIL error
>but if I pass it through the extras windows without applying any kind upscaler or other postprocessing tool and then click send to inpaint, it just werks

>>8690825
...is main broken for you?

>>8690829
I like experimental features.

>>8690830
then experiment.

>>8690843
would you like me to reinvent the wheel while I'm at it?

>>8690825
I updated when he returned for all the new snakeoil samplers. Haven't since as he was talking about dropping Reforge to branch off Forge or Classic. Point being I think there were only main, exp, and dev then.

>>8690511
change the seed, start genning, immediately stop and change it back
probably the fastest way

>>8690516
halving the gen time would be a 2x speedup, did you mean double?

>>8690789
>>8690792
if those are related its not generating anything anywhere near that

post what it made

>>8690886
garbage that no one wants to see

nongen or patreon-er seethe, call it

>>8690363
Fixed some inconsistencies, added some details. New image.

https://files.catbox.moe/ky79ho.png
https://files.catbox.moe/veix2v.png
https://files.catbox.moe/3souoz.png
https://files.catbox.moe/dwtzjh.png
https://files.catbox.moe/rdgvf1.png
https://files.catbox.moe/zdbsgs.png

slop

The slop reminded me that water is wet, but AI water is wetter.

any good resources out there for getting good at goon sloppa? comfyui specifically

>>8690917
zesty

>>8690930
Could you be more specific? Also I don't think the UI choice matters here.

I'm using civitai for making pics

this but novelai 4.5 only 25 a month

>>8690526
Thanks to our vigilant report submitters, this toddler got justly deleted! Yay.

>>8690968
i just wanted to know her age

>>8690948
I want to be able to consistently make art on this level of whatever character I want

>>8690433
>>8690789
>>8690713
>>8690186

>he slid /hdg/ off the catalog
lmao

>This thread is archived
oh no no no

Anyone figure out how to add multiple sections of text in various areas of the image on NovelAI?
>dialogue
>text on clothes
>text in background on posters etc.
Whenever I try to do more than one it gets all garbled. I just use the prompts "text, english text, text: (words go here)". Not sure how else to make it consistently create legible words without using the text: command, which it seems you can only use once per character without fucking it up

>>8691041
/aco/ has a rule where if you want to make a new "X thread" you have to post six on-topic images in it. I always wondered why.

why is he hating ayakon so much tho

>>8691046
/h/ has same rule. it's to prevent thinly veiled request thread spam where the guy essentially makes a thread for other to post content for him.

vvgggh it's going to be so hard and annoying to download a single image and paste an op over!!!

>>8691044
I don't think you can do this without separating parts of the image. you can inpaint separate sections and add text separately, that should work. pretty sure the model is capable enough to do what you want in one shot.

>>8691054
already back

>>8691057
based bibicii poster

>>8691052
Sure doesn't look like it from the rules post, and even if so I've never seen it enforced here.

>>8691088

>>8560405
>5.1) "itt [ass/bj/whatever] thread" and only posting one image is a request.

Doesn't explicitly say "6 images" I guess but it's still implied. I think it did mention six images at some point though, or maybe I'm confusing it with some different board

>>8691085
4x better = 4x threads
shrimple

>>8691096
ah cool, that looks like a valid report reason then

>>8691107
>report
Oh noes...

>>8691109
need that feeling of being in control
not like we can do anything else other than SUFFER

i love using loras for models not designed for them

Is there an actual good realism model? Every illustrious mix I have tried looks so plastic and unrealistic

>>8691261
>>8690090

>>8691261
Wan, chroma

>>8691267
do either of those do nsfw and can I train a lora so I can do porn of what looks like some 10/10 dressing as my waifu getting fucked

>>8691273
you can but you're probably a vramlet so no

>>8691280
I have 24gb vram, never tried training a chroma or wan lora though

neck biting is a lot harder than it should be

>>8691261
It's always a compromise between realism and anime knowledge. But most look alright as long as you don't prompt characters with anime bias.

Check the previous /aco/ realism thread for some discussion on it. My workflow is genning with SamBase, then low denoise img2img+controlnet with a non anime-based model like Lustify.

anyone got a fkey vpred lora?

>>8691336
I tried to do a vampire scene previously. The girl kept biting her own arm or some other body part EXCEPT the damn neck.

>>8691367
box? was it just rng? I had one where she bit his hand which was okay but it was always the guy doing the biting, especially if you do the neck bite e621 tag.

I know the hands are fucked. This is just an example.

>>8691370
We can't upload to catbox right now. I will post it as soon as their service is back up.

But yeah, it's just RNG. A solution would be to do a "arms around neck, nuzzling, hug" prompt, pray to God, then do some inpainting.

>>8691380
I mean it works in the end there's just no way to specify who is doing the action.

>>8690511
drag the gen over the workflow
there's some node pack (i think rgthree) that also lets you do this for individual nodes, i use this quite often with an image feed on the bottom

>>8691385
Ideally I'd just get AI to do the base composition and then photoshop and inpaint to finish the job but I can't even get that far. Vampire has like 15k pics and no one ever thought to tag neck biting for some reason. Crazy.

>>8691380
Neck/shoulder biting during sex is peak.

>>8691261
because realism illus model are finetune at best and lora merged at worst and none are actual model trained for it

>>8691394
>>8691380
>>8691367
SEX

>>8691433
Indeed, those pictures show sex. This is generally the kind of picture you'lls see posted on the /h/ - hentai board.

>>8691394
Reminds me about this nature factoid about how some species cause pain on their partners in order to signal when to release their semen/ovulate.

In your case of your cat girl, she is ready to get her litter of kittens.

>wanting to see a man's broad back as the center of your gen
very much the hallmark of homoeroticism me thinks

>>8691506
POVs get boring after a while. I want scenes. I want to see situations. I want more.

>>8691506
>lmao
>>8691512
I agree

>>8691512
Yeah, it's pretty hard to do anything really fun with POV, most of its value comes from "wow I get to see myself fucking my waifu!" which other people wouldn't really find that appealing.

pov you're getting a rimjob

Elven ears (erogenous).

>>8691564
please this is too hardcore take that elven degeneracy elsewhere

gonna make a nai thread hope youll join me over there!

>>8691564
Cut those disgusting long elf ears off.

>>8691583
kys ok?

>>8691506
In every image the character with the penis is me.

>>8691623
this except pegging

>>8691564
box?

>>8691640
he cant give you a box since he stole someone else gen and i2i it

>>8691646
I said give me the box.

>>8691646
real rich accusing anyone of stealing in an ai thread

>>8690750
catbox please

>>8691370
https://files.catbox.moe/fvwp8a.png

>>8691640
https://files.catbox.moe/kst6ws.png

>>8691646
It is indeed another anon's prompt though. Not i2i though.

>>8691679
How did you make those pics on the left?

>>8691682
Inpainting.

>>8691693
Obviously but what was the prompt?

>>8691696
>photo (object),no lineart,faceless,eyeless,photo inset,2girls,side-by-side,v,portrait,blue background,stelle (honkai star rail), march the 7th (honkai star rail),:D
Then I added some blur manually

>>8691708
Oh ok thank you.

>>8691708
NTA, but I was wondering how you accomplished that. That's quite ingenious. Will use it later.

>>8691655
it's artist:hiramedousa,

>>8690743
>and tell me how i should gen with noob on comfy?
braindead simple workflow attached

>>8692052
regular loras work best in 0.5-0.8 range

>>8692052
NTA but yeah this is unusable, huge slowdown before every genning step where gpu fans go silent then go vroom again after a while. (Checked VRAM, it's not full)

>>8692052
good stuff saar

>>8690973
>artist tags
>use a good model that doesn't fuck with your artist tags
>use tags as prompts
pretty much the basic

>>8692075
>(Checked VRAM, it's not full)
How much is free? I think if it's less than 6 gb, comfy won't try to cache the model.
Also genning at higher batch size should reduce overall time spent swapping the models.

>>8692150
Yeah it's less than 6. Oh well, no snakeoil for me.

>>8692052
how do I access the metadata? I thought 4chan strips all of that

>>8692154
>>8689703 (OP)
>4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcb

>>8692166
ty

>>8692183
box me up

>>8692214
Stealth meta data is right there

>>8692200
Huh upside pov is in fact not pov. Good idea for speed gooners like me.

it's always surprising to me just how few images can effect burn in, this is a difference of just 1.2% of the dataset

>>8692337
Wow I thought you were dead

He has realized that this is the superior thread and move on.

The time this guy has spent on a single lora is more than it takes for an average human being to learn to draw at a commercially viable level

>>8692337
can't exactly tell what's burned in from a single seed

looks like he brought a few schizos with him

>>8692357
how does one generate this level of sloppa

>>8692137
I would imagine it is a bit more complicated to generate stuff that is consistently high quality and keeping a similar aesthetic though no?

>>8692339
too busy with training to post

>>8692426
>how does one generate this level of sloppa
Do you mean that in a bad way or a good way? I just used a base prompt and a style lora with forge. I also made a mixed model of my own.

>>8692476
it looks really good man, I wouldnt even think this is AI if I didnt know the artist

>>8692426
mostly just artist tags mixed, little to no quality tags, little to no negatives and NAI-Vpred or some other models and the style should mostly stay the same or you play with the cfg

>>8692536
noted, I will try this out later.

>>8692539
i can throw you in a bunch of artist tags for you if you want or you look some up yourself

>>8692534
Thanks. I know some people dislike shiny skin stuff, and I'm prompting on an outdated GUI, but I just like the simplicity of it.

>>8692543
please do if you know any that look good

>>8692555
I think shiny skin stuff is pretty popular, isnt balecxi the most subscribed to AI artist?

>>8692579
>please do if you know any that look good

newest, [best quality, masterpiece, high quality:0.6], (pigeon666:0.75), (artist:muk \(monsieur\), tousen:0.5), (bigrbear:0.8), iuui, pottsness, BREAK, 1girl,miorine rembran,1boy,vaginal,on side,white leotard,leotard aside,highleg leotard,long sleeves,v neck,small breasts,covered nipples,:o,narrowed eyes,on side,fat man,white thighhighs,bed sheet,pillow,white theme,body hair,solo focus,size difference,pussy juice,dutch angle,grey background,looking to the side,sideways glance,leg hold,white choker,fake animal ears, cat ears,white headband,paw gloves,cat tail,neck bell,jingle bell,

>>8692536
made this with WAI and ran it through a face and pussy detailer, looks pretty good imo.

noob question: why does everyone seem to only use euler or euler_a? are they just better than the rest for anime pictures?

>>8692611
try khyle., possumachine, wamudraws, wagashi (dagashiya), fkey, honjou raita, zankuro,

>>8692614
WAI+Trendcraft lora, pussy detailer was fucking it up for some reason but this mix does look pretty good. A bit loli-ish due to some of the artists

I'll definitely be trying to mix artists around it seems to give a really good result overall

>>8692654
tokkyu,artist:chen_bin,artist:ask(askzy),artist:tidsean,artist:kasumi_(skchkko),artist:hiten_(hitenkei),artist:mika_pikazo, (artist:tianliang duohe fangdongye,morikura_en,misaka_12003-gou), lack, (artist:iuui,yunsang), (artist:ask_(askzy),rei_(sanbonzakyra),(artist:redi_(rasec_asdjh):0.6), (quasarcake,pottsness,healthyman:0.3), [sho_(sho_lwlw)],tianliang_duohe_fangdongye, as109]],(azuuru,torino aqua,(fuzichocono),keenh_da,yaegashi_nan,Kouji (Campus Life), (asanagi:0.8), (snegovski:0.9)|(free style (yohan1754):0.8)|(khyle:0.7}, (artist:roropull:0.25), (krekkov:0.25), (pigeon666:0.55), (ohisashiburi:0.75), (gogalking:0.55), (makkurourufu:0.75), [[artist:wlop]],[ningen_mame],[ciloranko],[sho_(sho_lwlw)],tianliang_duohe_fangdongye, as109]],(azuuru,torino aqua,(fuzichocono)

throw all into a mix with the rest of your artist tags and post result

hoooly

>>8692611
Not a huge difference but Euler tends to smooth things out, which is better for flat color and anime-style 3d renders. While the dpm line are more noisy, which tends to add more detail and is also better for realism. Photographs aren't supposed to be clean.

Also Karras doesn't work well with v-pred on a technical level, and some a11/forge people leave the scheduler on automatic which uses karras for the dpm sampler line. Meaning they will just look bad for them.

>>8692666
I've been using karras since I was told it is the highest quality one, which do you recommend?

Please disregard Satan's findings. He should redo his test just to be sure.

>>8692579
The hatred of skin shine comes mostly from how everyone else does it and how easy it is to achieve. Contrarianism perhaps, but it's a good thing. No other place will help you with avoiding the skin shine should you want to.

>>8690133
i'd appreciate if someone actually read the whole thing and shat and pissed on it.
https://rentry.org/yahbgtr
specifically looking for something i may have missed (and there's no todos) or explained poorly

>>8692673
Thanks for actually defining the terms. That will go a long way to helping everyone out. I'll give this a read too.

>>8692668
Maybe ignore the Karras bit. Seems I may have been misled by >>8651079 ages ago. At least from my quick tests rn, I do not see any negative effects when used on v-pred.

>>8692680
try 20 steps karras vs simple, unless it was fixed there'd be a huge difference

>>8692673
i don't think that that copier method guide that you're planning on incorporating is very good. it's from a few years ago and it's fundamentally wrong about stuff like needing to match poses and camera angles 1:1 or overfitting the input style lora being a good idea.

>>8692337
>almost 200 lora and this guy is still not happy with it
autism is truly a curse for some people

>>8692579
the most subscribed is artkoikoi

>>8692680
>>8692683
I asked grok and it did say theres merit to karras fucking up anime art with vpred models, recommended me to use normal. I'll do some tests to see the diff

>>8692660
just slapped all that on top of my artist tags dunno if thats what you wanted

>>8692684
i'm mostly planning to revisit it and show my own experiments, and maybe even compare it with other stuff like regularization images but yeah that guide is a bit iffy

>>8692536
>>8692591
Germanon giving genuine advice in the other thread. this thread dead in the meantime

why is everyone spamming my Miorine prompt now

its mine now nigga

>>8692696
who?

>>8692683
Not seeing a huge difference here.
25 steps https://litter.catbox.moe/8qh7hckgjd6aov2b.png
20 https://litter.catbox.moe/2pxipuswjc5365a4.png
15 https://litter.catbox.moe/fwpkydpvn0qj697v.png

seed:899504856129646
steps:20
cfg:5.0
euler_a
normal

>>8692713
karras

looks pretty much the same to me, I'm downloading nai v-pred 1.0 to test what happens

>>8692702
>Not seeing a huge difference here.
but a huge difference is clearly there, even at 25 steps. my verdict: you must be blind
>efficient nodes slop
>dpmpp_2m
anon...

>>8692687
it's not about being happy with it, i've trained versions of the lora which look good. i keep retraining it because it's my largest dataset with the most alternate outfits and i ported/made some of them myself. it's good for figuring out how shading can effect style burn in or testing out how large details on outfits need to be for the vae to work with them on top of figuring out hyperparameter settings between the two loras

>>8692726
I am open to suggestions, both on a better XY implementation and a better choice of dpm sampler.

>>8692736
there's no good xy implementation on comfy, the absolute best you can do is to craft each entry manually and make a header with a node
>dpm sampler
if you really want dpm one, it's dpmpp_sde
really though, euler/euler a is everything you need

>>8692740
Then I will continue using Efficiency Nodes for my XY charts. Why did you even comment on it if you know there's no better way?

>really though, euler/euler a is everything you need
We were just discussing the differences between euler and dpm before you showed up.

>>8692753
>Why did you even comment on it if you know there's no better way?
huh, dunno

https://civitai.com/models/833294/noobai-xl-nai-xl

seed:237726434088659
steps:20
cfg:5.0
euler_a
normal

all the same but karras

very... unique? v-pred seems to hate karras lmeow>>8692771

>hdg split back in may
>I completely missed it
oh shit, hopefully this is the place with more interesting tech discussion then
my current goal is to create some multi-gen stories, consistency is the main current question
I guess it simply has to be a lora for the characters?

>>8692775
Yes if you don't want to manually edit.

>>8692779
thanks anon, I will try it this weekend
which current training ui would you recommend?

>>8692781
ez-scripts

>>8692774
try adding sketch, watercolor or something along to the artist tags

>>8692728
would you mind posting the toml for one of the versions you ended up happy with?

dpmpp_3m_sde does some interesting stuff to v-pred


>>8692787
with karras still on?

>>8692787
well it sure tried to make something

>>8692794
you are not using a million artist tags now right?

Where is the anon that was promoting base noob? Where is he now that someone wants to avoid the "incorrect and, frankly, extremely dumb fucking behaviors" he was seething about? Retarded shitposters are never around when you need them, eh?

>>8692794
don't use wai, use NAI-XL vpred1.0 2d accelerated custom or any other cpred, WAI has a huge style bias messing with your artist tags

>>8692807
that was on noobai-xl-nai-xl v-pred v1.0 version same as this

>>8692804
not using a million artist tags on this

>>8692810
He means the one in the OP not base noob.

forgot the damn pic >>8692810

>>8692805
who are you talking about?

>>8692814
crtl + f the quote.

>>8692790
it probably isn't going to be very useful
https://litter.catbox.moe/uzhbw0a5mtdif9bw.json
https://litter.catbox.moe/gyt9mtpewbgotudt.json
i'm not sure how much the step count of the style lora really matters with this optimizer. i've had pretty decent success with this style lora approach when using around 2250 steps per style in it. resizing the style lora seems to help with preserving details some. so does reducing its weight. but ymmv will probably vary with both of those because the copier method has a nebulous goldilocks zone between the style lora being so strong that you need to train longer and the style lora being so weak that it starts burning in. if you're using multiple styles then including activation tokens can help sometimes if you have unbalanced concepts, i.e. all of your examples of outfit A come from koikatsu and all of your examples of outfit B are emu screenshots. but also that doesn't matter if you throw in a handful of 2d regularization images into the dataset and don't use activation tokens if you're using multiple styles.
at lowish character step counts that doesn't really matter though. with these settings i've been able to get around 800 steps when the training data includes one style and style tokens, around 1700 steps when using two styles and style tokens, and around 5000 steps with using two styles, no tokens and having 1% of the training images be slop regularization images.
all of that is also going to be dependent on the styles that you're using though, just koikatsu, koikatsu + xenoblade 3 and koikatsu + xenoblade x have all resulted in slightly or very different behaviors
i also train using loss masks and produce my own training data to keep composition relatively consistent between the style character datasets and the character training data and i'm not too sure how much of an impact that has.

>>8692810
keep cfg low at 3-5 and use the right clip skip

>>8692815
alright i found him, now what are you talking about?

sexy anime bitches itt

>>8692822
help the anon trying to use base noob to... use base noob...?

>>8692827
so, is this "anon trying to use base noob" in the same room with you right now?

>>8692816
all of that is still very useful as a reference, thank you!
>tokens can help sometimes if you have unbalanced concepts, i.e. all of your examples of outfit A come from koikatsu and all of your examples of outfit B are emu screenshots
that's actually something i've also had problems with before, so both of your suggestions (style tokens and regularization images) are good to know for helping with that

>>8692827
he's just someone fucking with samplers?
and while you should use the base model you shouldn't really raw-dog it without using a style lora(of something at least adjacent to what you're aiming toward) due to it likely fucking anatomy and wanting to explode when left to its own devices.

>>8692817
cfg5 clipskip 2, are you trying to help me use base noob? I was actually just testing stuff out not really trying to use it

>>8692839
everthing exactly the same but euler_a, v-pred just seems to hate dpm samplers in general

>>8692835
if you are going to use activation tokens then fwiw the ones you go with do make a difference. on base illustrious i go with koikatsu \(medium\) since it's already around 90% of the way there and 3d for whatever other style.
and if you are using koikatsu then make sure that whatever shaders you're using are consistent with the shaders used for training the style lora. this makes a huge difference. i had to use kkuts for some clothes and without manually adjusting the shading to mimic the studio shadows i had selected they looked insanely fucked up

any good prompts for detailed pussy spread?

>>8692871
detailed pussies are pretty artist dependent but tags like "clitoris" "urethra" and "cervix" can help

>>8692876
would you happen to know any good artist tags for that then?

>>8692881
try tenako_(mugu77) or laserflip

>>8692807
is custom better than final?

>>8692883
pretty good result but adding cervix keeps getting these weird fucked up close ups or inside shots so not that good of an idea.

>>8692890
yeah cervix might not work all the time because the tag is both for spread shots and x-ray shots

>>8692887
Yes

>>8692893
laserflip style is great but really likes to add in text even with negative tags against text

>>8692887
ja

>>8692901
Use negpip instead of negatives.

>>8692901
make a lora of his stuff with text edited out, probably using flux kontext to do it when necessary

>>8692911
Haven't learned how to make lora yet but I will keep that in mind

>>8692907
negpip seems to work rather well

>>8692930
thats a boy

>>8692183
>guidance_limiter_enabled: True, guidance_sigma_start: 5.42, guidance_sigma_end: 0.28, dynthres_enabled: True, dynthres_simple_mode: False, dynthres_mimic_scale: 3, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Constant, dynthres_mimic_scale_min: 0, dynthres_cfg_mode: Constant, dynthres_cfg_scale_min: 0, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: MEAN, dynthres_variability_measure: AD, dynthres_interpolate_phi: 1, latent_modifier_sharpness_enabled: True, latent_modifier_sharpness_multiplier: 0.3, latent_modifier_sharpness_method: gaussian, latent_modifier_affect_uncond: None
Why did he do it?

>>8692668
ER SDE + SGM Uniform is all you need, lil bro.

>>8692945
friend asked if I could gen some pics of his oc, how could I prompt black and white hair like this?

>>8692982
multicolored_hair, split-color_hair

>>8692939
All that is having toggled on sharpness in latent modifiers I turned everything else off;.

>>8692943
ER SDE was like twice as slow and wasn't even better the little bit I tried.

>>8692988
>twice as slow
Impossible. It's the same speed as Euler, Euler A, etc.
ED - Time taken: 6.1 sec.
Euler A - Time taken: 6.0 sec.
>wasn't even better
I mean, now that is subjective for sure but as a chronic sampler changer over the noob lifespan, it's the sampler I've finally stuck with since trying. Still, should not be any slower than the vanilla samplers.

>>8692989
no

whomstve is saying this?

>>8692994
Oh I apologize I thought it was the other one someone was shilling here a while ago, yeah I don't have that one on the reforge commit I'm on.

>>8693013
Oh so you're on an even more ancient pull that I was on before Pancho came back. I pulled when he came back and got it then but then some other anons said it had been added before that.
I think SEEDS 2 DP, 3 DP, etc were added when he returned and those are slow af.

>>8693013
Yeah that was me and dpm++sa is all you need. Idk what anon is talking about since ER SDE is shit and blurry. Maybe in img2img pass.

>>8693018
>dpm++sa
Only useful for i2i cn. Otherwise blurry and slow.

>>8692985
this got good results but I seem to be unable to control which side is white or black, any ideas?

>>8693020
It's slow but it's definitely not blurry, especially compared to euler of all things.

>>8693022
It's gacha or the simplest of inpaint jobs, anon.

>>8693025
I guess I'm off to learn inpaint

Anyone using grabber for image downloading for their datasets? It was working last time I used it maybe 2 weeks ago, loaded it up and no matter what I do it won't download any images or it will download like a handful and call it a day. It sees how many I filter for but only downloads the 6 for some reason and I can't see anything in settings that would cause this

>>8693037
Works on my machine and I used it a couple days ago. Try reinstalling.

>>8693037
Used it today and it worked fine, though danbooru was slow as fuck

keeps hallucinating my friend into a futa for some reason

you only get one of those

any secrets to tagging 2 different characters or do I just hit the gacha and pray to rng god it assigns the correct tags to the right character

>>8693066
On a 1girl, solo prompt? Are you tagging penises or something?
>>8693082
If the model knows them, you should be good. If it's a guy and girl, you can still pull it off with some perspectives and tag weights. 2 OC girls? Use forge couple.

>>8693100
supposed to be a boy, I was trying to pair him with another character but its giving wonky results.

>>8693107
oh yea it had futa tag there but it was still doing it without

going to yet another sex party

>dozens of bot posts reposting the same prompt
>futa spam
What the fuck is going on

how new are you

I have been staring at comfyui for the last 12 hours or so I think I have a problem

>>8693280
Why didn't you use one of those hours to inpaint, lilbro?

>>8693290
imma start doin that 2moro now my brain is fried, still learning.

>>8693294
Provided you're not shitposting, just a quick glance through that metadata nightmare shit that's comfy, get rid of that "lazypos" embedding. At most you'll use 'masterpiece, newest' with noob models provided they haven't been shitmixed badly.
Also, I get what you're trying to do with all the pussy prompts but it's best to make it detailed with inpaint. When you throw all those tags at your base prompt like that, you wind up with that gaping dragon you got there.
Good luck and keep at it.

>>8693299
question: how do I get backgrounds on 102d custom and r3mix? they really hate those... And why the hell does it like the foggy breath so much.

>>8693305
>backgrounds
You don't. At best you can prompt set pieces to fill in a scene but keep it simple and consider 'blurry background' to give the illusion shit's happening back there that isn't an AI thrown together nightmare.
>foggy breath
Probably a mainstay in the data of whatever artists you're prompting. It's just a bias many sexooo prompts will try to slip in there. Prompt each artist individually and see which one is responsible then adjust the weights.

>>8693311
WAI gets a lot of what I want right but I guess I'll need the inpainting for details.
Time for bed and i'll learn how to do it first thing tomorrow

>>8693022
train a lora

The holy grail of style baking
>perfect fidelity
>works on base noob without negs/nyalia/quality tags
>good anatomy
>no burnt in shine and thick linework
>capable of light and dark scenes even if the artist doesn't draw both
>coherent backgrounds
>doesn't gimp creativity
>small file size
Many have challenged the snakeoil dungeon to achieve this, but none have succeeded

bot post

>>8693371
>perfect fidelity
How do you measure that.
>works on base noob
Just train on base noob.
>without negs
You shouldn't be using them anyway.
>/nyalia
wut
>quality tags
Just use "newest, [best quality, masterpiece, high quality::0.6]"
>good anatomy
Unless you fucked up your training that's not an issue, and some artists just have bad anatomy.
>no burnt in shine and thick linework
That depends on the artist.
>coherent backgrounds
Retarded pajeet opinion, no porn artist actually draws backgrounds because they don't fucking matter.
>doesn't gimp creativity
Ok, that's the only legit concern so far
>small file size
Buy a 1TB SSD you bum

>>8693371
none of that is how any of this works you retard

what kind of reverse psychology is that

>>8693383
>and some artists just have bad anatomy.
There is also the fact that if you do have an artist that has bad/weird anatomy and your lora doesn't learn that then it's a failed bake and you have to go back to fucking around with training settings.

>>8693442
unless it goes to really weird/chibi proportions(read: something that's wildly different from the default) it's pretty much impossible to really get a lot of artist-specific anatomy to actually train without frying everything out.

>>8693324
Wai is well known for its good backgrounds so you'll either have to use a different model, prompt everything you want for the background, gen background first then inpaint characters in, or gen base res image in WAI and switch to 102d to change the style. I heard comfy chads can switch models in 5 secs so you should probably do the latter.

>>8693449
You'll get it if you train on illu 0.1 rather than base noob.

>>8693456
people train on anything else?

>>8693434
Man mochirong is a great artist.

>>8693456
No you won't. That's not how it works.
And training on illustrious 0.1 when you're using a noob derivative is actual snake-oil retarded garbage.

>>8693457
Some people were trying to push for a while to train on Noob or Noob Vpred. There is a slight benefit to training on a Vpred model for genning on Vpred, but the benefit really isn't enough to justify it when Illu 0.1 is pretty much the most "pure" SDXL anime tune we have.

>>8693461
I'm not gonna bother arguing with you since it's a waste of time.

>>8693462
"some people" meaning literally just about everyone but your retarded ass.

>>8693461
I'm going to make a statement from my own experience. I trained over 200 style loras on Noob and then I ended up retraining said style loras on Illustrious 0.1. I then ran comparisons on various models and shit mixes and the Illustrious loras were better every time.

>>8693459
Good for him, but that gen is just this >>8687795 lora.

>>8693465
Skill issue. Issue of skill.
When you're training a lora you're training a network on top of ANOTHER network. That network you're training is expecting the underlying network to work in really specific ways. If you change that underlying network, much of what you're lora is expecting to change will either not exist anymore or will function entirely differently.
In some cases the lora may in fact pick up and more or less replicate some bits of the underlying models network, in which case sure, it can actually "correct" where shitmixes get absolutely shitfucked. However it's still having to contend with all the other broken shit that it cannot fix. Also you have no control over any of this.

Which is all to say, if you're training on illustrious 0.1 and using it on anything other than illustrious 0.1, it's not going to be anywhere near optimal. Later noob models will have some of the underlying network in tact but the vast majority of it is not. This leads to a lot of issues like bad color output, general frying and bad anatomy. The only reason it wouldn't is if you're using a shitmix that has a predominate amount of illustrious 0.1 mixed into it.

These are the kind of stupid ass fucking opinions and """""conclusions""""" people come to when they rely entirely on shitmixes without understanding how any of it actually fucking works.

>>8693467
Oh ok thanks.

>>8693454
as in, gen on WAI then img2img on 102d?

>>8693479
Ja

Did you know that different models, even if it's a derivative of an earlier model, may in fact require different training settings and LR's for loras?
This means that settings used on illustrious 0.1 may not actually work optimally for noob eps 1.0! And that's not even getting into vpred, which requires a bunch of other stuff you actually have to change to even train it to begin with!
How interesting is that!

>>8693482
This is the least productive discussion I've ever encountered. The mark of a skilled person is making something complex seem easy and you are far from skilled. You could
>train a simple dataset on both models to prove your point
>post settings for both models to prove your point if you're to lazy to train
Instead we get mindless shitposting, the easiest thing any retard with working fingers can do, time after time after time. So fucking boring at this point.

Has anyone tried training on that bluevol model btw?

>>8693482
For those who can't actually read the underlying subtext here, this is to say that trying to A/B test a dataset against two different models while using identical settings may not actually work the way you think it will. Even worse, settings can change between datasets, so even if it works for one dataset it might not work correctly for another dataset. Even more worse, it's all fucking random and there are outliers where you could do everything perfectly identically, even down to the starting seed, but get two wildly different results.
Isn't machine learning fun? Are you having fun? I sure am having fun.

>>8693484
yeah man let me spend 15 hours training 4 loras on two models just to call you a fucking retard over something that should be common sense.
Or how about you learn to preface your "opinions" with "I use and this is how it seems to work" (which effectively translates to likely has a very large portion of as its base). Not only is that a lot more productive for everyone involved, it's infinitely more technically correct than omitting it.

>two options
>one painfully easier than the first that requires nothing more than an internet connection
>chooses the option with the greatest shitposting potential instead
I should have known. Total waste of time.

>n-no you're saying things contrary to me so that's just shitposting!
>s-s-s-so I'm going to shitpost too!
>look at me everyone! my shitposting totally invalidates the statements that I don't like!

did anything new happen with illustrious? They released 2.0 and then nothing else right

>>8693506
it's all gone SAAS as far as I can tell but also a failed SAAS that no one is using.

>>8693510
I can't wait for them to go belly up and then not release anything out of sheer spite, despite that they rugpulled the community not once but twice.

>>8693514
I'm not even mad since they gave us illustrious 0.1, and their later models can't even match noob.

>>8693054
based

>>8693107
>>8693109
NAI and other hentai finetunes of SDXL tend to be very "female biased", which means that unless you specifically prompt for male features it can be hard to get any actual 1boys, as opposed to a 1girl. And as you've seen, this female bias means you can get prompt contamination if you're trying to prompt feminine men. I've run into this issue a lot since it can be hard to get Asanagi-style "Big strong man roughly fucks a big titted girl", at least in my experience. You could try prompting "male focus" or "yaoi" in the positives and "1girl" in the negative prompt.

Also, be wary of directly posting stuff like this. I don't really care but /h/ has a "no gay shit" rule in general so you might catch a ban for posting yaoi.

>>8693604
would be dumb if a solo boy caught a ban, I avoided posting the actual gay sex I made him do.

>solo boy
>cock front and center taking up 60% of the image
>not /y/
Why is this place always infested with you people?

>Why is every place always infested with you people?
ftfy

>>8693634
Solo male counts as /y/ or /d/ but only if the dick is showing if you can't prove it's not a woman then it's okay

How do I get higher thighhighs? like almost the entire length of the leg

>>8693678
Scribble and inpaint.

>>8693678
Is there even such a thing? I think they end at the place they end because leg thickens higher up and if the sock ends at the thick place it would be constantly slipping towards thinner part. Basically, it wouldn't hold.

>>8693686
2b's are almost up to her ass in some pics

>boy
>actually massive futa cock
don't associate these fags with us gays
??

>futa
>dick and balls; no pussy
Wtf is a futa cock, anon?

iykyk

iiwii

>>8693694
is there a difference between male cock and futa cock?

>>8693686
all thighhighs need tape to stay up irl unless they're really tight but its not like thats a problem for anime characters

>>8693716
No. If you like cock, you like cock.

>>8693716
jesus do i need to do a futology class in here
what are you, 12?

>>8693716
>>8693721
Yes you dumb faggots. A futa cock is when the clit turns into a cock so they have two "futa" genitals, a cock and pussy. Doesn't make it any less gay.

who gives a shit hyper is /d/ and it's also aco as fuck

that's not the point bro
a futa cock is characteristic by the fucking swollen sack like balls and that weird tapered shaft
no faggot draws dicks like that because they don't look like that, it's a futa thing, it's even in memes

>>8693729
futa is also for fags thoughbeit

>>8693729
Anime website. If you don't even know Japanese you shouldn't speak about Japanese terms.

>>8693733
futa is for 40yo straight germans, same like le femboys

>>8693736
I don't care how deep they are in the closet, but they could at least have the decency to make their own thread

>>8693738
i assume the guy was just trolling considering he was posting solo futadick males in aco style here

Is there a good model to do genital inpainting for stuff like this or should I just keep trying with the made model\loras and messing around with prompts?

>>8693754
forget loras just use noobai

>>8693754
One of the most important things for inpainting in comfy is to set your crop factor really low (between 1.1-1.5) so you get a really high resolution mask. But if you set it too low, you start running into issues with visible seams and denoise being too high.

>>8693785
love the bliur

>>8693785
do you have a good workflow for inpaint? I got a very basic one from a dude on youtube but theres no such thing as crop factor anywhere in there unless thats reffering to the grow_mask_by thingy on the vae encode(for inpainting) node

>>8693729
never heard of this schizo theory

>>8693791
All the workflows I use are from here https://rentry.org/comfyui_guide_1girl#up-to-date-comfyui-guide-for-1girl-and-beyond

>>8693826
thanks that actually helps a lot, why is that not on the op directly...

>>8693832
Baker should add it next thread.

>>8693846
I think my stuff went up a level thanks to that link

>>8693888
That guide was what I was seething about last thread. /ldg/ is doing so well, has all the newfags, has all the guides, is actually trying to be productive, while this place is a shadow of its former self and filled with the same couple shitposters crying about the same five talking points.

I posted the wrong one like a retard

>>8693889
I should go there but I only ever gen porn so its hard to get feedback

>>8693889
> /ldg/ is doing so well, has all the newfags, has all the guides, is actually trying to be productive
the fuck are you talking about
that place has been a cesspit of low effort video spam for the past month or two.
it's dumbass boomerposting trying to be "le erotique wink wink" with shit that would be embarrassing to post to civit.

>>8693889
hey i'm trying to be productive here
https://rentry.org/yahbgtr

captain we need a bake

Testo testo

>>8693928
Welcome back, got any more Marnies?¿

>>8693933
No

>>8693936
Day ruined, I'm going to sleep.

Alright, hope I didn't miss anything

>>8693972
>>8693972
>>8693972

>>8693476
shaking in their boots when someone mentions empirical testing instead of imaginary armchair baking lmao

PAGE 11 BUMP