/jp/ - Otaku Culture

>>43564286
AI-ya is a cute

>>43564178
What AI service are you using anon?

>>43564294
Aya is always cute
>>43564300
stable diffusion

Drawsisters... it's owari-da....

>>43564752
you shouldn't say such things

>>43564324
Never used it before, what do I need to type in to get clean images like yours?

>>43565015
my prompt was sloppy as fuck anon, but here. also what model are you using?

Masterpiece, Best quality, maximum quality, best work, master artist, beautiful face, calm, smiling, lying on back, on bed, blush, half closed eyes, male pov, dramatic shadows, autumn forest, falling leaves, wind blowing, eye focus, detailed eyes, pointy ears, black hair, (red eyes), (short hair), (Shameimaru Aya:1.2), (black wings:1.3), (feather wings), (bird wings:1.1), (black feathers), kind, white button up shirt, skirt,
calm, smiling, sitting, sitting by a river, reaching out hand,
colorized, colorful, cinematic light, gold dust, dust, traditional art, film grain, dramatic shadows,
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, mutations, bad proportions, extra head, ugly, imperfections, disfigured, blur, grain, broom, white wings, panels, multiple views, painting, male, nsfw, bat wings, headwear, tokin hat, hat, teeth, pom pom /clothes/, red string, red hat,

>>43565060
damn that's a ton of tags
>what model are you using
by "I never used it before" I was refering to ai image generators in general

>>43565098
>ton of tags
I mean if you look at the prompt it as a lot of conflicting stuff in it. I was and am tired atm.
>I was refering to ai image generators in general
so for stable diffusion this is the guide you follow for nvidia gpus: https://rentry.org/voldy
Go to this site if and download a checkpoint that looks good to you: https://civitai.com/
if you want my checkpoints I'll throw them on mega for something for you.

>>43565133
Thanks, I've been thinking about getting into it for some time now, thanks for helping me out!
>if you want my checkpoints I'll throw them on mega for something for you.
that won't be necessary, there's more than enough on there, thanks

>>43565165
glad to help, also don't forget to set clip skip to 2.

Loli Aya with her shirt undone, huh.

>>43565060
>all those negative prompts
if you used an actually good model, you should barely have any

>>43567498
you sound like an easy negativefag, I see a problem in my gens repeat the first thing I go to is my negatives. if you think these are bad you should have seen old school schizo negatives.
https://rentry.org/dd5up4

also post gens.

>>43568216
It's too hot... GPUs are dying... Satorins are sweating... What is she even doing on the surface?

I am once again here to ask for your weird/reject AI gens, I love those for some reason

>>43568319
Well, all right. I've got this "Ringo" here, at least according to the prompt. I'm not entirely sure what's wrong with her, other than the obvious ears and the beret, but somehow it rubs me the wrong way...

>>43568387
Took the bunny tail a little too literally. But nah I'm talking stuff like this

>>43568387
I want her to rub me in all the right ways

>>43568319
Turns out Marisa was a jedi all this time

>>43565133
>nvidia gpus
I have amd…

>>43568488
https://rentry.org/sdg-link

>>43568488
amd requires more fuck fuck games, and totally bans you from touching text AIs

>>43568460
Pointy teef

Got a pretty good reimu after many failed attempts

>>43569446
I take it back

>>43570450
lovely

The Prince is wet.

Only so much I can do on 4GBs of VRAM.

>>43568488
AMD's shitty drivers are the reason I went with Nvidia, pretty much the only one in fact.
Don't know how it is for their CPU's though.

>>43572195
"Stop whoring around Yuyuko-sama"

>>43568399
In that case, this might count.

>>43572574
Yumeme LORA where?

>>43572195
>le AI can't draw hand- ACK!!!!!

this is the best i can generate
anyone know any better models i could switch to?

>>43573515
I'll let you in on a secret, since sure as fuck nobody bothered to tell me, and "beginner" tutorials tend not to mention it for whatever reason. Model doesn't mean jack shit beyond the basic art style and availability of elements (body parts, 2hu hats, etc.).

See, txt2img isn't supposed nor will it ever on its own produce "final" results. Sure, it's the "wow" feature of SD, but only a part of the toolset. What you're meant to do is generate a bunch of lower-res txt2img images as a kind of base. Then you pick the one you like the composition of, hit [Send to img2img], select "Resize and refill" and up the image resolution a bit, keeping in mind the correct proportions (this WILL eat up your VRAM, so you may get an error - just dial it down if you do). A plugin can automate setting the proportions for you.

Anyway, the option you should pay the most attention to is Denoising strength. It determines how heavily it'll base your re-generation on the provided base from txt2img. 0 is no changes, 1 is a completely new picture according to the prompt. I recommend 0.4-0.5, depending on how well you like your base. You can add or remove things from the prompt at this stage too, if you desire changes. Hell, you can edit the base image in photoshop and re-insert it into img2img with the changes while keeping the prompt.

This will basically take your "base" from a blurry mess to a proper, passable picture.

From there, let it shit out 5-10 results, pick the best-looking one, or merge them together in PS through masking if you like distinct elements but not the whole. And voila.

There are a bunch of things to learn beyond that, and it's a process, but if an idiot of my calibre was able to figure it out, so can you. Gun bar, eh?

>>43564752
i love ai art but the feeling of drawing your own wife with your own two hands is something that machine overlords can never take from you. art for money may be screwed (good riddance desu) but art for passion will never die

>>43573758
I managed to fuck up saving a .png. Don't be me, kids.
>>43573780
There's a middle ground. You can simply treat AI as a tool like any other and supplement your drawing with it. Have it fill in the colours if you aren't sure about that. Trace over gens. Draw low-fi, regenerate with a prompt. People love to doomsay, but SD is just a piece of software. It won't do all the work for you.

>>43573825
what prompts did you use for this image in particular?

>>43573934
>(hakurei reimu), 1girl, solo, (black hair, long hair), brown eyes, (red vest, red skirt), (yellow ascot), (hair bow, red bow), hair tubes, (detached sleeves, ribbon-trimmed sleeves),
>(gigantic breasts, hanging breasts, sagging breasts:1.3),
>(disgust), (gohei, holding gohei), hand on hip,
>full body, flying, from below, danmaku, sky, (traditional media:1.1), (masterpiece:1.1), (best quality:1.1),
The model turned out not to really know what a gohei or danmaku are, but the colours were cool, so I gave it the ol' CTRL+S.
>>43573515
Point in case, I took your picture and re-ran it through img2img with generic settings I homed in on. It's not perfect, but it could probably do with a few simple retouches in any editing software. Or a few masked parts from gens with a different seed. Etc.

>>43573758
You don't have to do the extra img2img step, just tick hi-res fix and it is done by default. It first generates a low-res image and then upscales it with whatever upscaler you choose.
I generally don't bother fixing any issues in my gens, just generate hundreds of pictures and usually a few of them will be decent. Quantity over quality.

>>43574017
>just tick hi-res fix
Hires fix is quite literally img2img. It's the same function. The difference is you get more fine-tune control with img2img, plus you don't waste time/processing power on upscaling/regenerating every single gen, most of which will likely be useless.

That said, if you prefer to fire and forget (and don't have to worry about the power bill...), it's a legitimate method too.

Nevertheless, for a newbie, it's better to play with img2img manually to get a feel for the settings and what they do. Then you can automate all you like and know with a degree of surety what's going to happen.

>>43573968
what do the numbers mean?

>>43574046
Tag's strength. A tag with no number is essentially strength 1.0. If you put it in parentheses and add a decimal after a :, you make the generator pay more attention to it. You can also bring the strength below 1 if you want it present but perhaps it overpowers other elements. You may have seen them denoted as multiple (((parentheses))) or {{brackets}}, but that denotation has since been phased out.

For example, since most Reimu pics in my model have her bust at... let's say, normal size, I had to up the strength of gigantic tit-related tags for them to overpower Reimu's. At 1.0, they were having basically no effect. Careful not to overdo it, as it can throw the prompt out of balance. 1.3 is already a LOT stronger than 1.0. They're not percentages.

Are any of you taking requests?

If so, please create an image of Star Sapphire hanging from a tree by her legs and her panties are showing

If you guys are taking requests then mine is to negate the request of the guy above me. Thank you for your time.

>accidently get SOUL in your path

>>43574070
That is a shockingly well generated reisen
If you made cleaned up some of the lines, particularly on the head, I wouldn't be able to tell at all.

>>43573758
I just generate big images and fix with inpaint and gimp. but sometimes I put two smaller image together in gimp and run it through img2img, then fix it in inpaint. So your method is fine, but it's very tiresome and not fun.

Are NovelAI posters allowed here?

>>43576316
of course, we're just here to take it easy and generate waifus

>>43576331
I only have a 3060 on my laptop so i'm not sure how good it'll do for AI

>>43576368
you'd be fine, I use to gen on a 1080 with only a little pain.

I love that since Touhou pics are so vast in numbers, you don't actually need LoRAs for popular 2hus like Reimu and Marisa. Models are trained on booru pics in general.

goocone Fr

>>43577377

bunny

>>43573825
>>43577894
Fucking great

>>43576261
The important part is to realise that raw txt2img at a low resolution (512x512 in the case I was replying to, geez) is only supposed to be the start of the process. Which is something I don't see mentioned often enough, and it skews expectations, including mine.

how do you guys get such detailed backgrounds?

>>43580543
It's mostly down to the model. You can try generating at higher resolutions and also try latent upscaling. I like prompting for things like "distant mountains, distant forest" but it might just be placebo.
You could try a model like this
>https://civitai.com/models/24779/dark-sushi-mix-mix
and see how that works out, I got cool backgrounds with it without even trying.

>>43575961
This model really likes Reisen in a nurse outfit and so does my dick.

>>43580901
What model?

I've been thinking about Gu's semi-canon underwear today.

>>43580963
Universal Solvent Mix.

>>43583055
Now show us what she's looking at!

>>43587517

>>43588163
Scary!

... But then why does she look aroused to see that?!

Seiran on Earth.

>>43589122
Don't use AOM3

>>43589331
No worries, I only used it for, like, 2-3 pics a few months back and never touched it again.

I'll treat this as an excuse to post another bun.

>>43588163

>>43589356

>>43589364

I have no idea how this thing works and i'm not interested in AI art but has anyone tried generating ZUNart using AI?

How do I generate something that doesn't look like it came straight from a nightmare?

>>43589433
just fuck your shit up man

>>43589374
>>43589364
>>43589356
Really god damn good LORA. Makes me feel like such a food for having generated a whole pornographic series with Akyuu using just tags and photoshop.
>>43589433
Learning to "feel" what the model understands and what it doesn't is a process. Make sure you aren't using conflicting tags. Make sure you're providing enough. Get a competent model for what you want to do instead of one of the generic, popular ones. Be aware that first-stage generations tend to have a big miss ratio.

Otherwise, several different flavours of advice have already been posted in the thread. Read up and train with the tools you were given. SD is just a program. It won't do the work for you.

>>43589533
>Really god damn good LOR
the only one that has a lora is the one with a sarashi, it it was a sarashi lora. I went full autist back in October on night while drunk off sake and made a prompt on novel ai that can every so often make gen a proper looking akyuu. two of those images are just novelai gens ran through a different model plus controlnet.
Sometimes I miss novelai.

>>43589533
Am I using the wrong app? Even when I'm copy pasting other people's prompts I still get these results.

>>43589593
I stand corrected. There is in fact an exceedingly good LORA out for A9 on Civitai. Thought you guys were taking it out on a spin.

If you're running a local instance, I can't recommend it enough. Attached is a result with no... OK, minimal cherry-picking using said LORA and a minimalistic prompt. It's amazing how well it works.
>>43589645
Ah. I take it you're using a remote service. Not familiar with those at all, sorry.

>>43589666
I tested it out after posting my gens, it's quite nice desu satan

Anyone know what model is this?

>>43589684
Human dot exe

Sad Google Colab doesn't work with this anymore. I gotta get a better GPU

>>43589716
If you're able to run SD at all and generate a basic picture in a non-excruciating time, then there have been optimisations released that might put off that buying decision. I generated >>43572195 on a laptop 3050Ti in a very comfortable time, for instance. Look up "multidiffusion upscaler" and use its "Tiled VAE" feature.

>>43589687
Nope it is AI, but the guy that was generating these didn't said what model he used.

>>43589794
I have an M1 MacBook with 8 gb of RAM. I made this on it. More than 512 x 512 images makes SD have a stroke.

>>43589830
You're running into the memory allocation error, right?

>>43589835
I don't know what that is desu. If there was an error it never showed up. Only thing that happens is the images end up being extremely blurry and distorted.

>>43590388
That face belongs on a Letty.

>>43590388
Sacchan LORA where?

>>43574017
>>43574030
catbox?

>>43593604
no way you AI generated a perfect Nekuro Marisa

>>43593740
It's the best model I've trained

>>43593764
Damn i wish i wasn't a total brainlet in this
I could get my own free Nekuro art

>>43593604
That's pretty good.

>>43591954
Okinaposter here. If you want the metadata, it won't help you much, since I engage in the forbidden technique of redrawing.

>>43591954
>https://files.catbox.moe/i5n29e.png
lora
>https://civitai.com/models/16512/usami-sumireko-touhou-project

She's Seki, but more importantly, she's Banki.

>>43593764
>>43593604
Please tell me there is some way to import that model.
Also, would it work on different characters?
Or is it only good at marisa?
Science must be conducted.

>>43597596
Love this one, could you post the uncensored pic as a catbox link?

>>43597752
Sure, here goes tits!

https://files.catbox.moe/bhi76r.jpg

>>43596788
I kneel.

>>43596994
Thanks!

>>43597977
Thanks, I like Seki's bankies.

Posting Horse

>>43599705
holy fuck, MOAR

>>43599705
Uncheck restoring faces, PLEASE

>>43599810
It's just one of my first attempts at using Adetailer, i still don't know how to use it all that well

>>43600269
Dear lord that's really good.
If there is an option to do so, a few more passes in the algorithm might sort out the oddities in the clothing, but damn that's a near-perfect Aya.
Even the eyes are pretty convincing, and normally that's how I rank AI images.

>>43575932
This healed my soul

>>43597624
>Also, would it work on different characters?
Yeah, picrel is with a LoRA for Shion but all the really popular 2hu characters will work just fine.
>Science must be conducted.
Here, please share your own results if you would like to (2gb model because I trained this before LoRA was even a thing): pixeldrain com/u/eRiHRfUR

>>43602253
I want this

I want this NOW

>>43604607
Cool, but it really would have taken you two minutes to fix the buttons.

Come to play in the underground?

Thanks for coming!
https://files.catbox.moe/rpfgvc.jpg

>>43605841
I love how the AI imagined lingerie consistent with her normal dress. also lol at the internal view

>>43605841
I have come to the conclusion that pantyhose >>>>> tights
I wish her hair was a little more interesting too, it seems like a lot of these models just go for the generic M shaped bangs
pretty good picture apart from that though

>>43572218
What's wrong with amd drivers?

teh yuya

>>43589178
Holy fuck those are some nice looking whaletits. Did you make the LoRA yourself?

>>43589178
can it not do the hook hair?

Dropping Birb

>>43607357
Nice, prompt and model plz?

>>43606387
Hmm, [forehead] gives somewhat over the top results (I think for obvious reasons), and I can't much feature Yams without the bangs, but you did give me a little bit of an idea to do something with her hair, so here's a twofer. Hope you enjoy fat spider ladies on the prowl!

https://files.catbox.moe/3qtpam.jpg

>>43607423
https://files.catbox.moe/adz0t7.png
Sure, I use AOM3-R34, i found it to be quite balanced

>>43607580
Can you make a Nue with thick thighs and spread legs like the birb in >>43607357 using the Nue LORA that dropped recently?

>>43607612
Forgot link https://civitai.com/models/77554/houjuu-nue-touhou

Thick AI 2hu thighs give me life

>>43607612
NTA, and not necessarily what you asked for, but I'm not doing anything useful with my evening at this stage. And more Nues never hurt anyone, right? Metaphorically speaking.

>>43607612
That Lora was pretty damn fried

>>43608294
Nice transparent panties!

>>43608294
Here's a tip. Pre-generate with the fried lora at moderate strength (0.75-0.85), then load that image into img2img and re-generate it at a higher resolution with the lora turned down to 0.15-0.25. It gets rid of nearly all artefacts, provided your input image is half-decent.

>>43608335
>>43608294
Point in case. Noisy pre-gen -> mostly clean re-gen.

The wings need fixing, yeah, but you should know how to do that at this point.

>>43608363
>>43608294
>>43608133
Impressive, moar!

>>43608335
>>43608363
The problem with fried LoRAs like this is that it usually means the composition as a whole has a ton of bias. It's less obvious since you are unable to see the dataset if you aren't the one who trained it, but if you could it would be very likely that an image exists in it with a very similar pose, composition, style, etc.
It's somewhat challenging to overcome currently because 1) most users have no idea what they're doing, tons of misinformation exists, and 2) the trainer most use does not yet support any sort of validation of loss to detect when overfitting begins. https://github.com/kohya-ss/sd-scripts/issues/193

>>43608601
In other words, it's easy to "accidentally" plagiarise? Some wouldn't call that a downside.
Though it is because I prefer redrawing to working with LoRAs.
>tons of misinformation exists
That has been my general impression of the community since I got into this thing. Wonder if it's just because the thing itself is new, or perhaps some kind of (un)subtle gatekeeping.

Also, nice style on that one. The wossname - oil paint effect - is that a LoRA or model thing?

>>43606628
useless for anything but gayms and supposedly buggy on windows, on linux they're really good but still suck dick for computing

>>43606762
Yeah.
>>43607278
No, I reckon you'd need to inpaint the hair in.

How do you make these AI images
I have no idea how

>>43572195
what gpu?

>>43608686
CPU's are fine i've heard, but their GPU's suck ass for doing literally anything but gayming.
If you want to do SD, you NEED Nvidia, since it's based around specific architechture.

>>43608693
One of the replies says, Anon.

>>43608685
>In other words, it's easy to "accidentally" plagiarise? Some wouldn't call that a downside.
Pretty much, and you're correct some wouldn't think that's bad, lol.
I think the misinformation is because most people have no fundamental understanding of how diffusion models work, let alone AI in general. Those that know this stuff really well and actually write papers and such usually don't have the time to help or sometimes even write good docs and instead are working on improving the tech even further.
>is that a LoRA or model thing?
LoRA is involved but most of the style is coming from the shuffle controlnet w/ starry night as the input img. Kinda old gen so I'm not sure if it can be replicated 100% but it should give a starting point. https://files.catbox.moe/8uxobk.png

>>43608750
Cynicism aside, yeah, that's probably it. The whole AI generation thing boomed way too fast, that's the problem.

... So I say, but here I am, having had no idea what it was all about just a month ago. Oh, well. Thanks for the input. I had no idea about CN's shuffle model beforehand. Yet another thing to learn on the list. Have a Raymoo in return.

Posting more barely SFW Birb

>>43608833
You can catbox any NSFW ones, I'm really liking that birb.

>>43608833
>>43607357
Post more huge tits Okuu, I need to fap before sleeping.

>>43608901
Spacejin Birb

>>43609011
That looks good, spacejin LORA where?

>>43609023
Here: https://huggingface.co/sylveriate/lora-styles/tree/main
Thanks to AI i can see how some 2hus would look if they were drawn by artists that usually don't draw 2hus, like this Sinensian Birb.

>>43609050
wait, wrong link. This is the one https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/eshGVCLY

>>43609050
Can you do that same birb with her arms up, showing her armpits?

>>43609050
>Thanks to AI i can see how some 2hus would look if they were drawn by artists that usually don't draw 2hus, like this Sinensian Birb.

It's crazy how they are so credible.

>>43609050
>Sinensian
lmao how come it didn't generate a huge futa cock?

>>43609342
I fixed that anon's mistake https://files.catbox.moe/visjea.png

>>43609060
there you go

>>43609472
Thanks! Keep the lewd Okuus coming.

>>43609472
>>43609011
>>43609050
>>43608833
These would be perfect if she had her wings. How could you forget them?

>>43590388
Please anon, share the Sari LoRA

>>43612623
https://mega.nz/folder/trojgYqZ#xQVVNxI7Sh4kW2uvY-tEyQ
Massive disclaimer, this didn't go through much testing and finetuning. Her clothing might vary between her canon design, and hemoglobin a1c's design.

>>43610645
i didn't forget them, it just that my Lora has a hard time distinguishing between her Wings and that small cape, so i chose to omit it the both.

>>43613345
Can't you just put cape as a negative prompt? I mean you did get rid of her normal clothes after all.

>>43614548
The problem is that the AI thinks the wings are the cape, and gets confused.
Not obvious necessarily, but I think that's what he meant.

>>43573780
I think that we're gonna see a lot more hybrid AI/human art in the near future. think AI-powered drawing assist tools being added to digital art programs, etc.

I think people will like that more than just genetic AI generated art (it looks good now but since it's literally zero effort wit will by saturation become generic) and artists will feel more valued knowing the human touch is still in the game, while being able to be way more productive by automating the boring and tedious stuff away

>>43608686
I like gayms though and I don't like how nvidia has a monopoly on everything.

"Come to play in the underground again!"

"(I'm totally pregnant, aren't I?)"

>>43605841
Jesus man, you truly are onto something here.

>>43615627
>>43615635
wtf how do you make different versions of the same pic without changing everything at once?

>>43616315
Inpainting, you select an area and it regenerates just that area with respect to the rest of the image.

>>43616425
Inpainting is a bit inaccurate and more than a bit of a hassle. It's good for introducing new elements, but not as much for "editing" parts of the picture without either messing up or not quite fitting with the surrounding area. At least on the model I work with.

What I do instead is I img2img a whole new picture/set with the new elements I want for the variation prompted in. Anime models are usually very good at changing expressions and hand gestures, even at low denoising strength. Then, I slap the two images into an image editing program and mask the layer with the "new" parts in with a blurry (feathered) brush. It's more accurate than inpainting, and - this is actually big for indecisive people like me - if the RNG gives me some new elements I like the look of, I can mask those in too while at it.

On the other hand, things like the slitted eyes were inpainted in, because the thick outline prevents artefacting at the edges. After all, black is black.

It's kind of like >>43614737 says. The whole thing works best if you don't treat it as the be-all-end-all and put in some actual, manual effort.

Have a Heca.

>>43616425
What are your inpainting parameters? When I try I always get a blurry mess.

>>43564324
>stable diffusion
Doesn't that cost money?

>>43616509
You're probably thinking of NovelAI's commercial model for Stable Diffusion. No, SD itself is open and free, and the vast swathe of available and preferred models are completely free as well. Check /sdg/ in /g/ for guides. No piracy needed. Although you can, of course, pirate NovelAI's model. Don't, though. It's only marginally useful for 2hus.

The cost is really hidden in hardware. Though that said, with the latest optimisations and a patient streak in you, you don't need to splurge on a 4090 right away. I've had decent fun running it on a 3050 laptop at work.

Moo.

>>43616541
Anon, you should know that any good anime model is built off the leaked NAI model. You aren't using Waifu Diffusion, are you?

>>43616561
naisu gen

>>43616584
So it isn't free, is that what you're trying to say?

>>43616718
not that anon but as long as you've got something akin to a 1060 6GB or better it's pretty much "free" with minimal effort(it'll work with a lower end GPU than that but it can get pretty sketchy pretty fast).
automatic1111 is open source and gets installed/updated via git.
you can download models/lora's from places like civit or huggingface without issues or having to use sketchy torrent sites/magnet links/xdcc bots/etc.
It's probably not quite "free" enough for stallman but you don't have to pay for it with anything other than the time it takes to set up and update(when you're so inclined).

>>43616792
NAI-derivative models fall under fair use as long as you don't use them commercially. So, you know, don't use it for work. Whatever your job would need to entail for that to come in handy.

>>43616603
If Utsuho got her way in touhou 11..

>>43616792
>(it'll work with a lower end GPU than that but it can get pretty sketchy pretty fast).

Why, what's the problem?

>>43616909
Lower vram than that is generally just a not fun experience and restricts you to pretty small resolutions and GPU's lower spec will have some really slow render times(and the 1060 6GB is already pretty slow as is). I think it can technically run on Maxwell but, again, probably won't be a great time.

I wonder if anyone would make a website for random people make some AI images online like some sanicfag did.

>>43619263
Moko culo....

>>43616961
Low VRAM can be circumvented through tiled VAE (which is piss easy to set up). The only real concern these days is whether your GPU can use this and other optimisations and whether it takes Mokou's lifespan to complete an iteration.

>>43564752
>chanta

>>43620325
Maybe I'm operating on slightly older info then. Either way, I was using 1060 6GB as a baseline since I was stuck using one for a while and, while it was tolerable, it was pretty much only just. Going anything lower feels like it would be complete pain, especially considering how much trial/error/iteration can go into the process.

>>43620803
It's all fairly recent, yep. Tiled VAE is a game-changer, though. That Mokou up there? That's 4GB of VRAM at work. It's not completely fail-proof (for example, that clump of black pixels on her shoulder is a tiled VAE artefact), but it's already pretty good and only going to get better. The days of 8GB+ requirements are thankfully over.

Now, raw processing/iteration speed is a tougher cookie. --xformers can speed some cards up by half, but compatibility's still not across the board for it. It's all very new, honestly.

>>43564178
>easy
DID SOMEBODY SAY EASY????

I have been generating nothing but Yamame porn these past few days. Libera me from this fat, blonde hell!

Something wrong with my Hecatia

>>43624358
BGM: Pure Hell ~ Whereabouts of the Pandemonic Planet

>>43624358
Yeah. There's no caption on the tee.

>>43623570
catbox a few, mah nigga

>>43623570
stop making her fat
she canonically barely eats given she attacks kanako in 17.5 specifically because she hasn't eaten in so long

>>43628965
I've posted WIPs in this thread before, but here's the finished, complete "set," I guess:

https://files.catbox.moe/ecnz9i.jpg
https://files.catbox.moe/65cfxq.jpg
https://files.catbox.moe/c5myd5.jpg
https://files.catbox.moe/khl6y4.jpg

And then the last two ones are: >>43615627 and >>43615635

I need to stop this... my spider sense isn't as much tingling anymore as it's throbbing angrily.

>>43630335
Well, I did perpetrate this set a while ago, before I figured out how to work with certain little things such as colour and higher resolutions. Perhaps it'll be more to your taste.
https://files.catbox.moe/i6yf59.jpg
https://files.catbox.moe/sem9xo.jpg
https://files.catbox.moe/iwfvqk.jpg

>>43630335
yam has been afflicted by the fat fetishist curse, much like patchy even though she's an anemic magician that doesn't eat at all

>>43630335
No, keep her fat!

>>43630373
Spiders have big abdomens, what's more to say?

I felt bad for letting a spider monopolise my time and dick, so I committed some busty ladybugs.

Full versions:
https://files.catbox.moe/6nyhuc.jpg
https://files.catbox.moe/duy2xv.jpg

>>43632248
what vae to you use

>>43632313
Just your bog-standard Anything_V4.

What's your take on the whole style "theft" debacle that's occurred recently?

>>43636282
lol
lmao

Aya that i made
I don't have all those fancy gadgets and my laptop only has 128mb vram so i use a modded apk of imagine AI to do all of this in my phone.

Very limited, fucked up hands and feet but it's good enough

Hope I can make it as good as you guys

>>43637267
You get bonus points for the unorthodox pose, but whatever deal with the devil you just described there, I don't want to have anything to do with it. My soul must remain pure.

I'm sorry.

https://files.catbox.moe/ybr7lk.jpg

>>43639141
What LORA is that?

>>43589835
Also if I could allocate more RAM for SD is there a way to do that using command prompt or something?

>>43640237
Your best bet is installing multidiffusion: https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111
Then install it (Extensions -> Install from URL), apply and reset, and lastly tick Enable Tiled VAE in this panel over yonder. If you're still running into memory allocation problems, lower the Encoder Tile Size (it'll default to an "optimal" value for your VRAM, but that may not be accurate) until it works. May give you a lower it/s, but it'll let you render at much higher resolutions. Also, remember you need to Enable it separately for txt2img/img2img.
>>43640219
Just Imminent Penetration from here: https://civitai.com/models/9432/pov-imminent-penetration-1-mb to prevent the dick going in. Which isn't a thing you'd want usually, but...

>>43640268
>>43640237
Brain fart. Just slap that url in the Extensions/Install from URL tab and hit Install, then click Apply and restart UI in the Installed tab. No need to "install" anything else externally.

>>43640268
>>43616541
Is it true that proompting too much rapes your power bill?

>>43640278
As much as running your PC at high load for long periods of time, playing vidya at an uncapped framerate, what have you. It depends entirely on your individual PC. Got a 3090/4090 and are fire-and-forgetting batches of 100 via Hires fix? Well, that's tantamount to running your electric kettle for an hour straight.

I exaggerate, but it's very much possible to overdo it. It's incomparably more efficient to render at lower res then upscale selectively, but that, as some in the thread have said, can strike you as tedious. It's your call in the end. You figure out your optimal workflow.

>>43639976
Love those colours. What model?

>>43640268
>>43640274
Okay, I installed it. What do I do from here? Sorry if I'm being retarded, but I genuinely don't know what any of these buttons or sliders do. If you were me, what would you enable, disable, etc? Obviously I would want the highest quality and highest resolution images I could feasibly get on my machine without memory issues, so increasing the "encoder tile size" and "decoder tile size" is supposed to help with that?

>>43640601
use one of the Karras++ samplers, not Euler a(ncestral)

>>43640618
To add to that, don't use Euler A(ncestral) because it tends to mix early (less defined) stages of the generation into the final result. It's nice for purely creative or abstract imagery, but that's presumably not what you want. Samplers are, superficially, snake oil that won't matter for a beginner, but you should steer away from samplers with S in the name - it stands for "Slow" and doubles your Sampling steps regardless of the slider. Meaning it takes more time to gen. Or just switch to DPM++ 2M Karras if you want my take. It's one of the faster ones and the Karras thingamajig does... beneficial things, as I understand.

Optimal CFG scale to start is 12-15, if you're being specific in your prompt. Sampling steps have diminishing returns over 30 (although that may vary per model). You should LOWER Encoder Tile Size if you're getting memory problems, as that divides the picture into more, smaller "tiles" for your GPU to render one by one.

You can use either Hires fix (to mass-upscale all results) or img2img (to upscale only your chosen one) to iterate on your result. Denoising strength means basically how "new" the re-generated picture is going to be. 0 is leave everything as-is, 1 is completely new. Generally, I don't leave the 0.4-0.5 range for that, to keep the existing composition.

If you img2img, each jump up in resolution is going to add more detail to your pic. So it's better to go through each pic 2-3 times than upscale straight up to your target res. Pay attention to keep the aspect ratio of your source picture. There's an extension that can do that for you.

Txt2img is never, by itself, going to produce acceptable or final results. Get used to that. It's not meant to be the end of the process. Either Hires fix or img2img your results for proper effects.

That's all off the top of my head. Experiment, check each option out at extreme ends to see what it does, figure out your golden middle.

Have a 2hu nobody likes. She needs a home.

>>43640668
Can the AI make a decent Hina yet?

>>43640482
https://huggingface.co/autismanon/modeldump/blob/main/mokoumixv1_prunedfp16.safetensors

>>43640697
I'll give it a crack on Sunday if nobody else does in the meantime.

In case anybody wants to copy the settings for img2img, here're mine. How it goes is I'll usually generate a couple dozen results in txt2img at 600x480, pick one I like, load it up into img2img, bump the resolution up to 1000x800 and fix the prompt (if needed), re-generate 10-20 pictures at 0.45 Denoising strength, pick the best one and then repeat the process at 2000x1600 / 0.4 Denoising. That gets me the best results without killing my 3070. Regenerating twice also yields pretty fine details in the final pictures.

Also, keep in mind you can always redraw parts of the gen at any stage in any editor and pass it through img2img for convincing results at... eh, a fairly minimal artistic skill requirement. That's the secret sauce and the real killer feature of SD. So if you get an extra arm, or need to get rid of one of Sagume's wings (wink, wink), you can crudely do that to your initial, low-res txt2img gen - before loading it up to img2img and pretending you're just THAT good at prompting.

Extra tip: (Resize and refill) is better than (Just resize), as it fills the empty pixels in with "noise" that the AI can interpret all anew, rather than the chonky pixels of your upscaled picture.

Are there any CBT/groin kick LORAs yet? asking for a friend of course haha

Western Okina

>>43646000
Realistic teeth on anime girls is something I didn't know I needed until now

>>43646000
Turn off restoring faces, PLEASE.

>>43640668
I'm nobody. Share the uncensored version on catbox please

>>43646000
Looks good

>>43640697
I generated these a few months back with a hastily trained dreambooth model. It had some difficulties getting her hair right and also kept turning the stitching on her dress into hair.
>https://catbox.moe/c/ddkmt9

>>43646897
Aye, captain:
https://files.catbox.moe/c9qhu8.jpg
And, since I'm easily provoked into unsolicited continuity, here's two follow-ups:
https://files.catbox.moe/yfy3c9.jpg
https://files.catbox.moe/2u5ag9.jpg
Nobody asked for it.

>>43650588
more hina
>https://catbox.moe/c/p7neqq

>>43640697
I had to cheat just a little bit to get all the design elements I wanted in in one go (Hina isn't Hina without her boobie slope somewhere in the picture), but otherwise that was pretty easy. No LORA.
>>43654678
I'll clue you in on something, because it baffled me for the longest time. To get rid of the "bits of sleeve hanging off the arm" problem, just... specify the sleeve length in the prompt. Otherwise the model will draw on both short and long sleeves at semi-random, resulting in disconnected bits such as is dangling from Hina's forearm there. It's the stupidest thing.

Bonus Hina.

Oopsie doodles, I completely messed up the straps that time.

Here goes again. Bonus Hina doll in a babydoll.

>>43655627
make pregnant Hina
the best kind of Hina
putting 0.6-0.8 weight usually works well

>>43655952
You seem knowledgeable. Why don't you do it?

>>43654678
>>43650588
>>43654824
Thanks, these look really good!

Hecazuri

can someone prompt some hot ai images of ushiromiya rosa

I was going to get an XPS13 laptop to replace my ancient stinkpad but this thread is making me strongly consider getting a gayming laptop instead, if only to restore the pit miko's dignity by generating lots and lots of non-cowtit images.

Do you guys think that anything would be possible with a GTX1050ti with 4GB of VRAM and a 4th gen i5 or would it just be a waste of time? Because that's what my desktop has.

>>43663657
Definitely possible but it's obviously gonna be slow, probably about a minute per image

>>43662423
can you just assume that every time you post a censored picture, I want you to post the uncensored one too? It'd save me a bunch of posts

>>43663990
Hey, it gets you to reply, which gives me an incentive to post.

Here's an uncensored version:
https://files.catbox.moe/9v9vd9.jpg
And here's an apology Junkozuri:
https://files.catbox.moe/i47s0v.jpg

>>43651157
>>43601173
Vanilla marisafag best poster itt

Somehow, a satisfying Hecazuri keeps eluding me.

https://files.catbox.moe/jwgusw.jpg

a gappy from back in january

>>43664351
>junkozuri
What a slutty tattoo.

>>43671711
>>43671794
DO SOME REIMUS DO SOME REIMUS

>>43657639
cute

I've committed a series where the playable mainline characters are looking at you while wearing skimpy bikinis.

... Also, their pubes are sticking out, for some reason.

https://files.catbox.moe/cytvxa.jpg
https://files.catbox.moe/fypizz.jpg
https://files.catbox.moe/t26zr7.jpg
https://files.catbox.moe/8r2par.jpg
https://files.catbox.moe/dakgs9.jpg
https://files.catbox.moe/mt89s5.jpg

>>43680268
what is this, fucking vietnam?

>>43680285
Worse. It's an actual Asian country.

I forgot these two honorary playables.

https://files.catbox.moe/u88f82.jpg
https://files.catbox.moe/is1xfv.jpg

>>43680268
Reimu is best due to her expression

>>43686965
>Cirno
>Pubes
Ruined

>>43688690
That Aya is neat though.
I'd like her to have it a little more whispy, and slightly larger boobs, but other than that it's good.

>>43688835
meant for >>43688790
I fat fingered it after fucking up.
Also Sakuya looks the best out of all of them.

>>43688835
I too take my Aya slightly on the bigger side (which looks great in a dress shirt), buuut I had a theme (boner) going, so yeah. The optimal Aya would be something like this:
https://files.catbox.moe/2029y6.jpg
>>43688790
THEME
>>43688783
>Reimu is best [...]
Agreed.

>>43680268
That does it, there's no point in drawing anymore. *smashes tablet and throws drawing supplies*

>>43689715
Said this before. I do, actually, do a bunch of drawing and redrawing when need arises. With actual drawing skills and some basic prompting, you could do a damn sight more than I'm capable of. SD is just another tool, albeit a revolutionarily accessible one, granted.

Youmu is in danger.

https://files.catbox.moe/lc8yul.jpg

>>43689589
I just realized you're posting these on Pixiv because I'm also downloading from aibooru and that Aya one showed up in my dup filter in Hydrus.
Are the ones you're posting to Pixiv the cleaned up but also censored versions? I can't actually tell because the image as a whole is slightly different between the two, and the one you posted here is also slightly higher res, lol. Sorry for the dumb instrusive question but just curious to hear what your workflow is.

>>43697463
Oh, sorry about that. I'm still VERY new to all this and picking up new things almost daily. That Aya up on Pixiv was posted while I still hadn't been aware of some things. The one up here in this thread is just a regenerated version of it I grabbed as an example of the optimal Ayasize, since it was the only Aya (with her titseimarus out) I had lying around.

The versions up on Pixiv are sometimes "updated," as I keep tinkering with details between postings due to overweening pedantry. Censored, too, of course, because Pixiv's really anal about that. No way to reupload/update without Premium either.

Uh, don't judge me for the obscene descriptions. Just trying to give the pieces I put up SOME merit over the regular AI spam.

>>43697647
Ah, that makes sense then if it was regenerated, no worries. Happy to see the uncensored versions being posted here tho, keep up the good work.
I won't judge your descriptions either because I'm doing some similar bullshit where I have a LLM generate captions for my own enjoyment :^)

>>43697723
Speaking of things that are sadly going to have to be censored, I've played more with this pose today.

Miko:
https://files.catbox.moe/nkbgzw.jpg
https://files.catbox.moe/gxpgpm.jpg

Heca:
https://files.catbox.moe/ec1oos.jpg

Reimu:
https://files.catbox.moe/cxykr5.jpg
https://files.catbox.moe/r4kfvf.jpg

>>43698525
Nice. Reimu is best as always.

>>43698525
Damn I love those cumshots, just drew a white line then inpainted I assume?

>>43698606
Not quite. Scribbled a rough rope of mayo on the initial, 600x480 pre-gen, then img2img'd it up twice to target resolution with appropriate things added to the prompt. Then masked the relevant part of the new gen onto the old pic in GIMP (which is how I get these practically identical variations).

A bit more work, yes, but inpainting is a crapshoot with this model unless it's with a specific LORA. Can never get the edges to blend well enough. With my workflow, it's easier and faster to draw things manually than try to RNG the inpaint.

>>43698577
For speaking the universal truth, you get one more.

https://files.catbox.moe/m9cl0o.jpg

>>43698980
Can you do a Nue one pls?

>>43698986
Might be a tall order with that fried LORA, but I'll see what I can pull off either tomorrow or on Sunday. Same type of pose, I take it?

>>43699091
Yeah same pose if possible, with the thickest thighs with skindentation you can get away with. Honestly still not sure how that Nue LORA is bad, the results look fine?

>>43698656
Ah okay, that's not far off from what I imagined the workflow being, thanks for the deets. You have the right idea to paint it in at a low res first because when you upscale then you're giving SD attention to the whole image during inference.
Semi-related, you might find this interesting if you haven't seen it. Basically the ControlNet author assumes Adobe is using a cascaded pipeline for inpainting and built in a somewhat easy way to automate that. The focus of his post was mostly for prompt-less inpainting but it works super well when you prompt stuff too. It's how I inpainted the cock in >>43609463, lol.
https://github.com/Mikubill/sd-webui-controlnet/discussions/1464

>>43662423
How the fuck are those planets so good