[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip / qa] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/h/ - Hentai


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


Old >>7009679

>NAI LEAK SPEEDRUN+FAQ
https://rentry.org/nai-speedrun | https://rentry.org/sdg_FAQ

>NEWS + UPDATES
https://rentry.org/sdupdates3
https://rentry.org/sdgoldmine

>MIX RECIPES
https://rentry.org/hdgrecipes

FREE:
>Local install
Nvidia GPU: https://rentry.org/voldy | https://github.com/AbdBarho/stable-diffusion-webui-docker
CPU: http://rentry.org/cputard
>AMD Guides
Native: https://rentry.org/sd-nativeisekaitoo
Docker: https://rentry.org/sdamd
Onnx: https://rentry.org/ayymd-stable-diffustion-v1_4-guide

>SaaS
NSFW txt2img: discord.gg/unstablediffusion | https://pornpen.ai

PAID:
Hentai: https://novelai.net/
>Cloud Hosted Install
Colab: https://rentry.org/sd-colab-automatic
Paperspace: https://rentry.org/865dy

MISC:
>Models
https://rentry.org/sdmodels
https://cyberes.github.io/stable-diffusion-models/

>Textual Inversion
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Textual-Inversion | https://rentry.org/hypernetwork4dumdums | https://i.imgur.com/yv3TrrC.jpeg

>Resources
Artist styles: https://rentry.org/anime_and_titties | https://rentry.org/artists_sd-v1-4 | https://files.catbox.moe/9wgqx9.jpg
Big index: https://pharmapsychotic.com/tools.html
Tags: https://danbooru.donmai.us/wiki_pages/tag_groups
Tag effects: https://pastebin.com/GurXf9a4
First pass calculator: https://preyx.github.io/sd-scale-calc/

>Booru
https://aibooru.online/

>SaaS Sites
txt2img: https://huggingface.co/spaces/stabilityai/stable-diffusion | https://beta.dreamstudio.ai
img2img: https://huggingface.co/spaces/huggingface/diffuse-the-rest
inpainting: https://huggingface.co/spaces/fffiloni/stable-diffusion-inpainting | https://inpainter.vercel.app/paint

>Other Guides
Inpainting/Outpainting: https://rentry.org/drfar | https://rentry.org/inpainting-guide-SD
Upscaling: https://rentry.org/sdupscale
Animation: https://rentry.org/AnimAnon
Wildcards: https://rentry.org/NAIwildcards

>Related
>>>/aco/sdg
>>>/g/sdg
>>>/d/ddg
>>>/vg/aids
>>>/e/sdg
>>
>>7011207
based Rinposter
>>
File: 99.jpg (2.16 MB, 2048x3072)
2.16 MB
2.16 MB JPG
>>
File: tmp6a7hb5xg.png (1.63 MB, 960x1408)
1.63 MB
1.63 MB PNG
I hate model mixing gacha
>>
>>7011215
Rin is cute
>>
File: 2.png (360 KB, 512x512)
360 KB
360 KB PNG
Know what I don't see around here too often? Transparent clothes
>>
>>7011220
I've been doing a trinart/f222/NAI merge gacha for a while.. Idk it all feels like it's coming out like shit.
>>
Daily wife

https://files.catbox.moe/x358fy.png
>>
Prompt suggestions for her?
>>
File: 12.png (366 KB, 512x512)
366 KB
366 KB PNG
like this
>>
File: 98As5xyFinal.png (1.96 MB, 960x1408)
1.96 MB
1.96 MB PNG
>>7011223
I know that feel. There is this chinese message board somewhere that shares like really nice mixes and models among themslves, but it requires a Chinese phone number and a account to get into, and sadly I'm not a chink so back to model gacha for me.
>>
>>7011225
seggs XDDDD
I need to get the AI to make a fem spurdo for these threads
>>
>>7011221
very true
>>
File: dead.png (2.44 MB, 1216x1600)
2.44 MB
2.44 MB PNG
>finally get the perfect pic
>hand is backwards
>can't inpaint for the shit of me
I read the guid but It always comes out a blurry mess for me, can someone give a QRD on how to inpaint?
>>
File: PriestessAI.png (2.31 MB, 1024x1536)
2.31 MB
2.31 MB PNG
>>7010750
Thanks a LOT to whoever did this. Looks very interesting.
>>
>>7011180
>nsfw, masterpiece, best quality, (red), (1girl, orange eyes, sitting on intricate burning throne, blue fire fireball, with oni horns and ponytail hair, hair over one eye, wind, serious, shiny hair, perfect face),strap,golden samurai armor, armored high heels, medium breasts, (highly detailed), shrine, scroll, spot color, field of blades, planted sword, detailed pupils, lips, intricate ruins background, pillar , (dramatic lighting)
>Negative prompt: deformed, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, malformed hands, out of focus, big boobs, long neck, long body, child, nudity, furry, fuzzy, signature, watermark, (anime), 1boy
>Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 2608919420, Size: 1280x1280, Model hash: 19810fe6, Denoising strength: 0.65, Clip skip: 2, ENSD: 31337, First pass size: 0x0
>>
>>7011233
inpaint at full resolution
prompt hand and background, or even just use the same prompt you txt2img
leave steps/cfg the exact same
set denoise to around .65-.75
set resolution to 1024x1024

and let it rip
>>
>>7011234
nice timing, I literally just sat back down at the computer.
>>
>>7011217
if only the titty didn't merge with the arm
>>7011226
good idea, will try it
>>
>>7011233
>>7011240
This is definitely one way to do it. But honestly you'll get pretty far by just rotating it in you preferred image editor first. That should at least allow you to get more consistent results.
>>
>>7011241
Do you think it possible to 'fix' the hands? If only a bit?
>>
>>7011246
I could but at the same time I could pose an even better question, is that the image you want me to fix or would you prehaps like me to see what else I could prompt?
>>
File: 37.png (354 KB, 512x512)
354 KB
354 KB PNG
>>7011244
Bless ya
>>
>>7011248
As much as I would like to see more prompts, humbly believe its better to work on that one already and fix it a bit and then, if you are willing to do it, prompt more to see what comes out.
>>
Can AI draw footjobs
>>
So when is this stuff gonna start looking different from one post to the next? It looks great but its the same pose same face every time. Asking for a coomer friend who has already exhausted his AI kink, of course.
>>
>>7011258
just prompt a different artist
>>
>>7011258
The sameface is a NAI feature. There's literally only one face it makes unless you put in lots of facial feature details, and seemingly no one does. It makes images generated with NAI trivially easy to spot
>>
>>7011255
see I'm of the opposite camp, yeah I could refine a image but I could also just prompt another 100 in that timeframe and find something better

>>7011261
oh hey you removed all the words proud of you
>>
>>7011257
Not without some work on your part. Anon has tried with hypernetworks and dreambooths and it always comes down to number of toes and ankle anatomy. Someone should fund a full blown model training specifically for feet stuff, maybe some day the chinks will.
>>
>>7011263
I'm not the guy who wanted it done. I did it in 20 minutes 2 days ago and forgot to post it lmao. Just deleted my post because the image is pretty offtopic and I uploaded it accidentally
>>
File: Untitled.jpg (1.85 MB, 4098x1536)
1.85 MB
1.85 MB JPG
>>7011255
oh sorry I didn't attach an image but heres a couple, I'm generating some more but if you say have position or theme or any preference I could add it. want some gobbos? tentacles? facelessmale?
>>
>>7011259
>>7011262
Do you guys touch it up in photoshop or another drawing program? Or is it all the program? I’d think an artist who starts using this would be cranking out really good stuff if they learned it
>>
>>7011263
>see I'm of the opposite camp, yeah I could refine a image but I could also just prompt another 100 in that timeframe and find something better
That pic is spot on for me (minus the hands) so thats the reason of the attachment.
>>
>>7011270
promptu?
>>
>>7011273
i dont touch up things because it's kind of tedious and the fun part for me is making prompts, not perfecting images

i did make a short manga type thing but that's just speech bubbles and even then it took a few hours to write the dialogue and get it all set up >>7003307
>>
>>7011273
Personally, almost everything I do is inpainting. I can't into GIMP to save my fucking life. The images are a big hit when I post them to my dA though.
>>
File: tmpkfv763jl.png (1.53 MB, 1024x1536)
1.53 MB
1.53 MB PNG
>>7011273
i use GIMP and inpainting on mistakes that bug me too much. otherwise its 90% the program.
>>
>>7011270
pretty
>>
>>7011270
>want some gobbos? tentacles? facelessmale?
Nope.
>I'm generating some more but if you say have position or theme or any preference I could add it
You are already SPOT ON
>>
>>7011277
(extremely detailed CG unity 8k wallpaper), (masterpiece), (best quality), (ultra-detailed), (best illustration),(best shadow), (an extremely delicate and beautiful), finely detail, classic, (manga).
nsfw, a completely (nude) tall loli with long blonde hair and blue eyes leaning on a wall in a dark cave. full body, scared, she is using her hands to cover her privates. small breasts, thighs, big anime eyes
negative is the schizo prompt
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7.5, Seed: 2525575375, Size: 1024x1536, Model hash: 6569e224, Denoising strength: 0.57, Clip skip: 2, ENSD: 31337, First pass size: 512x768
>>
>>
>>7011207
can someone please make shirai kuroko hentai?
>>
File: 27.png (348 KB, 512x512)
348 KB
348 KB PNG
Love them gals who're just happy to be here, aye?
>>
>>7011289
The monkey paw delivers
>>
>>7011278
>>7011279
>>7011280
Have any of the artists who you’re prompting noticed? Not like they could do much because I don’t think there’s copyright on porn drawings. Like what if you started prompting Kubos style to make some Bleach girls look just like his I wonder.
>>
>>7011300
well styles aren't copyrighted so nothing anon
however if you took his character and drew it in his style and then sold it that would be different now wouldn't it, just like any other copyright thats existed for years
>>
>>7011300
noticed what that i am using their name to make the ai draw cool lighting? i'm not replicating stuff 1:1
>>
>>7011300
I don't usually use any artists in my prompts. And how the fuck would they notice me, I'm a tiny insignificant speck in the deluge of generated garbage being posted continuously.
>>
File: tmp6ufglg11.png (1.19 MB, 768x1280)
1.19 MB
1.19 MB PNG
>>7011207
does anyone know what this model is b0a9077?
>>
>>7011301
>>7011302
>>7011304
I usually just scroll through the images in these kinds of threads so I was just wondering what the creators were thinking is all. Keep it up
>>
is https://rentry.org/sdmodels no longer maintained, or is there really no new public models?
>>
Hi guys, I finished my in-depth tutorial on using Stable Diffusion: https://rentry.org/sdhypertextbook

Tell me if there's anything that I'm missing/that you'd like to see
>>
>>7011311
no worries, answering questions is part of conversation
>>
File: 03888-20221125134529.png (778 KB, 640x960)
778 KB
778 KB PNG
Anyone know how to get a good fellatio gesture without summoning disconnected cocks :(
>>
Repostin

>>7011113
masterpiece, best quality, 1girl, braid, (perfect_tits), oval glasses, long_hair, lips, navel, nipples, vagina, solo, (black_hair), elf_ears, brown_eyes, (pubic hair), black_eyebrows, teacher, uncensored, nude, eye reflection, smooth skin, glossy skin, classroom in background, 4k, (big eyes), (large_breasts:1.2), (pointy breasts:1.3), ((by_zettai_shoujo)), 40 years old, mature female, (wide hips), serious, hands down
Negative prompt: (((monochrome))), 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, (animal ears) , (holding eyewear)
Steps: 50, Sampler: Euler a, CFG scale: 10, Seed: 3443717638, Size: 832x1216, Model hash: 0d70159d, Denoising strength: 0.69, Clip skip: 2, ENSD: 31337, First pass size: 0x0
>>7011120
masterpiece, highest quality, blurry background, rim_lighting, facing viewer, small breasts, nipples, (yorha no. 2 type b), short silver hair, black blindfold, full body, (sketch:1.3), traditional media, monochrome, (black and white:1.2), portrait,
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, futanari, deformed_hands, (((realistic photo))),
Steps: 24, Sampler: Euler a, CFG scale: 9, Seed: 2926277997, Size: 1024x1024, Model hash: 0d70159d, Denoising strength: 0.69, Clip skip: 2, ENSD: 31337, First pass size: 0x0

In general, combination of (sketch) and (monochrome) with negative prompt (((realistic photo))) will get you a nice sketch-like drawing.
>>
File: tmpmd390kxw.png (1.1 MB, 640x1152)
1.1 MB
1.1 MB PNG
>>
>>7011325
I got decent blowjob entirely by luck
>>
File: 00427-2974375717.png (2.3 MB, 1024x1536)
2.3 MB
2.3 MB PNG
>>7011275
>>
>>7011325
describing the man a bit more seems to help a bit.
>>
>>7011337
PERFECT Thanks a lot.
>>
File: 03887-20221125134241.png (746 KB, 640x960)
746 KB
746 KB PNG
>>7011338
>>7011334
>>7011325
The gesture is specifically the hand over the mouth, kinda like this, but nai really wants to add a lil bean penis.
>>
File: 02099-20221114230717.png (488 KB, 576x576)
488 KB
488 KB PNG
>>7011342
holy shit lmao, someone call then feds
>>
>>7011344
give prompt, i'll try
>>
>>7011342
Does she look like one tho?
>>
>>7011344
ah crap i missread, try putting penis in negatives if you haven't already.
>>
>>7011316
What's the prompt for these look like?
>>
>>7011351
in a word, unreliable

https://files.catbox.moe/ebmbvl.png
>>
File: 00033-2582129080.png (503 KB, 512x768)
503 KB
503 KB PNG
>>
File: 00811-~1.png (907 KB, 768x1024)
907 KB
907 KB PNG
>>7011307

Haha, I recognize her. Mostly.
>>
>>7011357
Whoa nelly that's a powerful image.
Prompt me up please
>>
>>7011334
>>7011342 no anon, get it right.

>(((realistic)))
>photo
>10yo child
>>
>>7011367
indeed
>>
File: xy_grid-3080- .png (3.95 MB, 3456x768)
3.95 MB
3.95 MB PNG
I love REISEN!
>>
anyone know what model is Model hash: 81f21c5e?
>>
>>7011373
Prompts or catbox please?
>>
What's the reason pedos post here and not on /b/? Wouldn't it be nicer to have all the content you like in one place?
From what I've briefly checked even /d/ has less of them than this thread does
>>
>>7011378
>have all the content you like in one place?
but anon, it already is.
>>
File: 00005-3792942776.png (2.54 MB, 1024x1536)
2.54 MB
2.54 MB PNG
>>7011329
thanks anon I retooled it a little
>>
File: 20221120204209- .png (666 KB, 576x768)
666 KB
666 KB PNG
>>7011377

Original Prompt is not mine. I just did the sequence to test if my sequence is working for any prompt. I find that it does work for many prompts.

https://files.catbox.moe/noszkc.png

https://aibooru.online/posts/9224?q=reisen_udongein_inaba+

Script: Plot X/Y, X-Axis, Prompt S/R:

"sfw and she is facing the viewer
","
nsfw, lying, on back, she is having (sex:1.4), (pov:1.5), legs_spread, (pussy juice trail:1.2), embarrassed, saliva,( penis on pussy:1.170), hand on own chest
","
nsfw, (vaginal sex from behind:1.45), back focus, (pov:1.2), (from above:0.5), (cum in pussy:1.0), (pussy juice trail:1.2)
","
nsfw, (pov:1.2), pov crotch, she is giving (fellatio:1.6) to a man's (cock:1.2), saliva, deep kiss, close-up, eye focus
","
she has a mouth full of cum, solo focus, pov, close up, open mouth, (cum on tongue:1.1), cum_in_mouth
","
nsfw, she is facing the viewer angry and (cum is dripping from her pussy:1.2)
"
>>
>>7011299
seems fine to me
>>
>>7011382
Ohh, interesting. Alrighty, thanks a lot anon!
>>
>>7011299
I just want to ravage that flat chested lesbian's pussy
>>
>>7010528
>>7010669
>>7010603
>>7010304
>>7010224

Please, share prompts!
>>
Not the kind of swinging I had in mind, but I'll take it.
>>
File: xy_grid-3082- .png (3.87 MB, 3456x768)
3.87 MB
3.87 MB PNG
>>7011382

This sequence has a high success rate. Rolled 6/6 image coherent.
>>
File: xy_grid-3093- .png (3.29 MB, 3456x768)
3.29 MB
3.29 MB PNG
>>7011289
I tried.
>>
File: 1654281626196.png (1.46 MB, 768x1280)
1.46 MB
1.46 MB PNG
Kek what's going on here
Someone caption this
>>
>>7011411
(bad anatomy:1.2)
>>
File: ddimryza0201_comparison1.jpg (1.23 MB, 5632x869)
1.23 MB
1.23 MB JPG
learning rate - 0.00000172
steps - 10800

10800 feels like it's been overfitted, very difficult to make it do things outside of training set. it does make very accurate faces, so it might be useful for inpainting
i feel like the goldilocks zone is somewhere around the 8400, but it's hard to say

what learning rates / step counts you guys use for your dreambooth models?
>>
>>7011412
KEK
>>
I was gonna experiment tonight but this prompt was so good it got me horny
>>
File: xy_grid-3097- .jpg (1.96 MB, 3456x768)
1.96 MB
1.96 MB JPG
>>7011382

Around elves, watch yourselves.
>>
File: Gridley_202211242133128.jpg (1.57 MB, 2364x1576)
1.57 MB
1.57 MB JPG
Date night for breast-conscious Blonde Elf Knight-kun. Apparently leave allowance is short.
>>
>>
>>7011443
I really like this style. What's the model/prompt?
>>
File: 03908-20221125151608.png (586 KB, 576x832)
586 KB
586 KB PNG
>>7011344
ayoooo
>>
I lost my Satou Kuuki hypernetwork and didn't find it on the links.

Can someone repost it here please?
>>
File: tmpmavugzak.png (1.66 MB, 1024x1536)
1.66 MB
1.66 MB PNG
>>
Anons help me out, what are your best fingering prompts? I can't get that stupid shit going no matter what I try.
>>
>>7011456
oh the tasty little slut anon is back!

ever walked in on her in the toilet or shower?
>>
>>7011465
masturbation, hand in pussy, inpaint the fingers
>>
File: yourmom.png (477 KB, 512x512)
477 KB
477 KB PNG
There is something kinda special about low step number (1000 to 2000 with 1e-5) overblown networks like gelu 1,4.5,4.5,4.5,1. They are big enough to take time to start converging, but they also quickly get a general grasp of every single detail of what you show them and they still haven't sorted out a common denominator between what they have inside them.

What I am telling you anons is this is how you would go about making your own artstyle if you want to peddle stuff on pixiv. Make a huge network and cut it off well before overcooking or converging happens.
>>
>>
File: yourmom2.png (289 KB, 512x512)
289 KB
289 KB PNG
>>7011472
I mean this is nowhere close to what I want it to do but it does look nice.
>>
>>
>>
>>7011486
idk why but this gives me nostalgia for when I first started prompting.
>>
>>
File: file.png (648 KB, 512x768)
648 KB
648 KB PNG
>>7011480
>>7011486
>>7011490
Cute!
>>
>>7011493
huh? I just some said some shit about nostalgia...
>>
File: tmpfo7sgpa1.png (1.56 MB, 1024x1536)
1.56 MB
1.56 MB PNG
>>7011468
what a coincidence...
;)
>>
>>7011494
I meant to ping click on >>7011491 but I have fucking retardation! hehe xd
>>
>>7011493
Thanks!
>>
File: tmpfeod8kov.png (624 KB, 768x512)
624 KB
624 KB PNG
When the LSD kicks in.
>>
I changed my gtx 960 to a rtx 3060 and for some reason, all of my prompts wont work the same way as before.

It is because of the gpu change? Is there anything I can do?

Pic is the old image generated at the gtx 960, next post will be the "new version"

((masterpiece)), (((best quality))), nsfw, 1girl, (face focus), black long hair, (black eyes), (spiral pupils), (hypnosis), (mind control), (brainwash), ((corruption))
Negative prompt: lowres, bad anatomy, bad hands, (poorly drawn hands), (poorly drawn face), poorly drawn legs, error, text, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, monochrome, dark skin,
Steps: 30, Sampler: Euler a, CFG scale: 20, Seed: 1697935505, Size: 512x768, Model hash: 925997e9, Clip skip: 2, ENSD: 31337
>>
>>7011500
This is the new one. Same prompt, checked already.

Before I was using --medvram, and I'm still using it. Already tested without it, and the result is the same.

((masterpiece)), (((best quality))), nsfw, 1girl, (face focus), black long hair, (black eyes), (spiral pupils), (hypnosis), (mind control), (brainwash), ((corruption))
Negative prompt: lowres, bad anatomy, bad hands, (poorly drawn hands), (poorly drawn face), poorly drawn legs, error, text, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, monochrome, dark skin,
Steps: 30, Sampler: Euler a, CFG scale: 20, Seed: 1697935505, Size: 512x768, Model hash: 925997e9, Clip skip: 2, ENSD: 31337
>>
>>7011495
You're a lucky anon to have found her in the latent space
>>
I found that the WD 1.4 tagger seem to give slightly better results than deepdanbooru, anyone knows whats the pipeline to use it for embedding?

from a lazy fag's point of view,
do I do a batch tag first, then slice it in preprocess images (auto focal + split oversized images)

or do I do that in reverse (preprocess then batch tag) ?
-
>>
>>7011234
what's your prompt for the priestess. Seems pretty cute
>>
File: kotoyoshi-hyper-72500.png (1.72 MB, 1024x1536)
1.72 MB
1.72 MB PNG
Mogubro here, Kotoyoshi Yumisuke hypernetwork is ready. This one was also a bit of a tricky one, I guess cause Kotoyoshi's art style has changed so much over the years. I had to cut down a lot of the oldest and newest ones to stay in the sort of 2004-2018 years where the changes were less prominent.

https://mega nz/folder/hlZAwara#wgLPMSb4lbo7TKyCI1TGvQ

More pics and training info in the folder
>>
>>7011510
I do the reverse myself, that way you aren't getting tags for things that suddenly aren't there anymore
>>
>>7011331
Hot. Can you share the prompt for this?
>>
>>7011456
Don’t have to give the full prompt, but what tags are you using for this babe’s clothes?
>>
File: tmplncdsvhb.png (1.73 MB, 1024x1536)
1.73 MB
1.73 MB PNG
>>7011468
w-what are you staring at, anon?

>>7011547
i don't mind.

realistic:0.8, (photo realistic:0.65), (extremely detailed CG unity 8k wallpaper:1.2), cinematic lighting, (full body), detailed shadows, 1girl, petite, short:1.5, chibi:0.3,(personal ami:1.3), __haircoloring__, ass visible through thighs, (dark brown hair), bob cut, messy hair, blue eyes, black latex thigh highs, black eyeshadow, black mascara, black eye shadow, bags under eyes, exposed breasts, cleft of venus,latex elbow gloves,latex fingerless gloves,exposed pussy,aroused, (flat chest_1.5), (small breasts_1.7), (small chest_1.5), long eyebrows, __danbooruhair__, navel, collar, nipples, nude, parted lips, (skinny:0.8), (athletic_0.6), pussy, smug grin, solo, stomach,
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, pink eyeshadow, pink eye shadow,
Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 12, Size: 1024x1536, Model hash: 12dd7ee3, Hypernet: j_k, Hypernet strength: 0.842, Denoising strength: 0.57, Clip skip: 2, ENSD: 31337, First pass size: 512x768
>>
>How do I keep the color of things correct in the prompt?
I'm having a little syntax issue where the color of things is incorrect:

nsfw, (masterpiece:1.3),best quality, (highres:1.0), (high quality:1.3), 4k, detailed, intricate, soft lighting, (realistic:1.2). (perfect face:0.9), (dynamic lighting:1.2), detailed, facing viewer, from below, full body, prostitution, mature woman squatting in dark alley at night, (shiny dark purple mini dress:1.2), (large breasts), large ass, (sanpaku:1.3), leopard print thong, cameltoe, skindentation, glowing red eyes, silver long hair, hime cut, aroused, steaming body, sexually suggestive, spoken heart,
https://files.catbox.moe/7h5coy.png
https://files.catbox.moe/wrtd40.png
>Prompt returned rel pic.

I asked earlier about framing prompts and got a decent answer, but now I'm wondering how to make the thong lepoard print and keep the dress purple?
>>
>>7011550
>AI does the right thing seconds after I ask
https://files.catbox.moe/0up1mf.png
>>
>>7011549
uhh, what's the prompt for this?
I'm interested in that hand position.
Different anon from the other prompt asker
>>
Was there a cloud hosted dreambooth service that was paid, or was it simply a self deployed dreambooth cloud instance that people were talking about?

my 2070 can't do dreambooth locally it seems
>>
>>7011549
Which hypernetwork is this?
>>
>>7011561
masterpiece, (high res_1.2),(best quality_1.2),(ultra-detailed_1.2), (extremely detailed CG unity 8k wallpaper:1.2), (illustration_0.4), sketch:0.45,realistic:0.8, photo realistic:0.6,cinematic lighting,raytracing, shadows,nsfw,from below:1.2,detailed nipples,sharp focus, short hair, bob cut,blue eyes,naked,dark brown hair, small breasts:0.8, small chest:0.8, flat chest:1.2, messy hair_1.3, sitting, arms covering breasts, exposed pussy, detailed pussy, aroused, embarrassed, surprised, angry eyes, blushing, narrowed eyes, yandere eyes,sitting on a toilet, sitting on toilet, detailed bathroom,
Negative prompt: lowres, bad anatomy, missing arms, missing hands, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, monochrome, mask, fat, shit, scat, futa, simple background, chibi, hands, extra pussy, more than one girl, more than 1girl,
Steps: 48, Sampler: DPM++ 2M Karras, CFG scale: 12, Seed: 556733594, Size: 1024x1536, Model hash: 12dd7ee3, Hypernet: j_k, Hypernet strength: 0.842, Denoising strength: 0.56, Clip skip: 2, ENSD: 31337, First pass size: 512x768

>>7011564
j.k. (jdotkdot)
>>
My system was compromised, interesting.
>>
>>7011570
ah, you didn't prompt for her to cover herself like that, it is sad day!
>>
File: zamn.png (1.71 MB, 1472x1472)
1.71 MB
1.71 MB PNG
>>
>>7011572
What did it comprise of?
You did check your pickles before using them, right?
>>
File: dkasa4221K1l.png (1.65 MB, 1024x1536)
1.65 MB
1.65 MB PNG
>>
>>7011573
the 'arms covering breasts' might be prompting like 'arms squeezing breasts' which is making a large number of the pics i am genning from that batch to be in that pose.
maybe try those two and see if that gets what you're after. but so far, that prompt still makes 4 or 5 / 6 pictures doing that pose lol.
>>
>>7011574
Thats a nice clean whole you've made. I'm assuming there more witchcraft involved than just prompting it?
>>7011572
>>7011575
>system was compromised
Is this a windows thing?
>>
>>7011500
>>7011502
Try using
>--precision full
as a command line argument. Your new GPU would be using autocast instead of full. This will change your seeds, but full is notably slower than autocast.

This is normally the culprit if you're trying to match up images made with an older GPU or a CPU, both of which require full precision.
>>
>>7011579
Honestly this txt2img output was like 80% of the way there already
>>
>>7011572
wa happun anon
>>
>>7011572
can confirm, just got a ransomware attack using the chinkai v3 checkpoint. thank god i've been doing everything in a vm
>>
>>7011601
>he didn't check for pickles
>>
>>7011601
Good thing I only ran that shit on colab.
>>
>>7011608
Doesnt webui have one built in / does it automatically?
>>
>>7011608
>>7011610
What? Where is this built-in pickle checker?
>>
>>7011601
was it the full or the pruned models
>>
>>7011611
I've been using https://github.com/zxix/stable-diffusion-pickle-scanner
I have scanned everything I download and it's all been clean.
I've also been running regular security scans and my pc has been good so far.
But who knows, maybe I've been pickled.
>>
Gonna either say this is bullshit, or I’m about to get pickled
>>
HELP, I used chinkai and now my dog has aids??
>>
>>7011601
>>7011572

imma call bullshit on this.
niggas didn't post details + new pickle inspector from sdupdates show all models as safe
>>
>>7011608
sorry it was me originally. While I don't entirely believe it was anything v3, as I definitely downloaded something else today and I haven't had any issues till now so I believe I've already isolated the concern, it was a bit of a surprise as I don't think I've had something like that happen in like 20 years. I got lazy
>>
>>7011625
What else did you download, anon…
>>
It seems like fat/chubby or short body doesn't exist in SD
>>
>>7011625
If you learn what it is for sure, let us know.
I'd rather not use chinkai or any mixes containing it if that really was the culprit.
>>
>>7011630
>ccp file
Like I wouldn't trust it just in general, but no one has actually made any reasonable claims that it is pickled. I've certainly seen no evidence of it
>>
>>7011618
don't worry it's a condiment for asians
>>
>>7011634
Its just that anon being retarded by his own admission, but if its indeed a sleeping bomb, I think more than a handful of us will get fucked together at least
>>
>>7011634
/g/ analyzed it.

They concluded the full model COULD be pickled but the f32 and f16 models were clean
>>
>>7011644
I would if I had any idea how to make good poses or genitalia that didn't look like it was melted or fused into something horrible.
>>
>>7011644
prompt me a working mating press thank you
>>
>>7011642
So that solves it?
I ran the scan on my models directory and everything passed. When I ran it on my quarantine directory it wasn't able to complete, but thats likely due to a megadl script trickle downloadings someone's full mega folder.
>>
>>7011651
if you have useful information aboot pickled models share it here.

If it passes it passes
>>
>>7011578
I was talking about the hands being between the thighs like that
>>
Have a last minute Thanksgiving image, inspired by the anon who was captioning their images.
>>
File: 00062-1788674490.png (694 KB, 512x1024)
694 KB
694 KB PNG
>>
File: tmpcj89y7vq.png (885 KB, 832x832)
885 KB
885 KB PNG
>>
File: happy thanksgiving.png (2.23 MB, 1472x1472)
2.23 MB
2.23 MB PNG
Happy Thanksgiving
How do you like your turkey served
>>
>>7011720
Put mashed potato on her face and a wine bottle up her ass
>>
murakami bro, did you manage to finish your hypernetwork? can you drop a link when ur done
>>
File: .jpg (640 KB, 832x1088)
640 KB
640 KB JPG
Ka-ka
>>
why tf are my images so faded when using strawberry mix every single image is like this
>>
File: 00056-1065524046.png (1.53 MB, 1024x1408)
1.53 MB
1.53 MB PNG
>>
File: tmpt1d57dc6.png (1.39 MB, 960x1280)
1.39 MB
1.39 MB PNG
>>7011730
>>
>>7011734
Does she have a USB port in her asshole?
>>
>>7011730
some anon figured this out last thread: if you have set your VAE to "auto" it might not be loading. set it manually and see if that fixes it
>>
>>7011735
shes adorable.
prompt?
also I am getting those purple spots too just now, wtf.
>>
>>7011735
Dem purple spots.
Call the ambulance!
>>
>>7011739
see >>7011738
>>
>>7011225
cover her in cum
>>
>>7011738
yep that fixed it thx
>>
>>7011739
(masterpiece:1.3), highres, best quality, official art,1girl, upper body, petite, cute, long hair, brown hair, blue eyes, (camisole),:d,:happy,(bright) sun lighting
Negative prompt: multiple girls, lowres, text, error, missing fingers, (bad anatomy), extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, extra fingers, mutated hands, (multiple penises), (disembodied), (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad proportions), (extra limbs),(wings)
Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 3773373277, Size: 960x1280, Model hash: e043dfc5, Batch size: 2, Batch pos: 0, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, First pass size: 512x768
>>
File: 6-min.png (1.74 MB, 1792x2560)
1.74 MB
1.74 MB PNG
Inpainting is fun.
>>
Made with Novel Ai. Haven't tried diffusion but haven't yet been disappointed by novel. I may DL it soon.
>>
Can the gradio-img2img-tool color-sketch
be used with inpainting?

Looks like it's only supported in img2img itself at the moment.
>>
>>7011762
>>>/aco/sdg
>>
File: tn3324sa230s0.png (1.64 MB, 1024x1536)
1.64 MB
1.64 MB PNG
>>
File: uhh.png (2.23 MB, 1472x1472)
2.23 MB
2.23 MB PNG
>>7011723
>>7011720
>>
File: 423ak7das7d8.png (1.85 MB, 1024x1536)
1.85 MB
1.85 MB PNG
>>
>>7011720
Wait, so live turkey goes in and that comes out?!
>>
>>7011750
did you inpaint the hearts on pupils?
>>
>>7011779
>>7011720
You could probably make a pretty good joke out of this with an existing character that has fire powers.
>>
>>7011781
Yeah. The AI can do heart-shaped pupils pretty efficiently, but I didn't prompt for it in original image since it was part of a 'progression' series of images.
>>
Also ended up with this output which I thought was pretty funny and you can technically post on blue boards
>>
>>7011786
they look really nice. How much emphasis do you put on a new prompt when inpainting it? Do you modify the old prompt too?
>>
>>7011789
2 girls, 1 cup, you know what to do
>>
>>
>>7011797
very cute, anon. could we have the prompt please?
>>
>>
>>7011799
I didn't save the prompt I just modified one of the prompts found here. I think I was using Strawberry mix and Euler A
>>
>>7011802
dropping your pic into PNGinfo in WebUI gives you the prompt
>>
>>7011802
https://rentry.org/hdgrecipes
>>
>>7011803
Is there a way I can inspect the images here? My prompt btw was
(Masterpiece:1.2), (best quality:1.2), extremely detailed, absurdres (soft lighting), (intricate:1.2), (cowboy shot:1.3), nsfw, skindentation, big blue eyes, eyelashes, skinny, solo focus, long brown hair, (outside a beautiful detailed cosy sunny beach:1.2), sitting under umbrella, tiny breasts, (micro micro bikini:1.2), large thighs, sweating profusely, looking at viewer
Negative prompt: bad drawn eyes, deformed iris, simple iris, no pupils, heterochromia, 2D, monochrome, (((bad shading))), ((lowres)), (((bad anatomy))), (((bad proportions))), (((bad hands))), (((multiple arms))), (((multiple legs))), (((multiple feet))), (simple background:1.6), (jpeg artifacts:1.4),( signature:1.4), (watermark:1.4), (username:1.4), blurry, multiple views, censored, text, error, blurry, worst quality, low quality, normal quality, missing fingers, extra fingers, extra digit, extra limb, extra nipple, extra breast, missing limb, floating limbs, disconnected limbs,malformed hands, long neck, long body, fewer digits, deformed, disfigured, amputee, mutated, giant head, double bellybutton,
Steps: 28, Sampler: Euler a, CFG scale: 14, Seed: 2410613051, Size: 960x1280, Model hash: e043dfc5, Batch size: 2, Batch pos: 1, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, First pass size: 0x0
>>
>>7011811
thanks!
>Is there a way I can inspect the images here?
4chan strips metadata from images we upload here, unfortunately
>>
>>7011381
looks nice
what's prompt did you use?
>>
>>7011730
>>7011735
>>7011738
>>7011745
Holy shit, I had no idea about this. Tested it myself, confirmed that's the solution. I've updated the hdgrecipes rentry with some example pics and explanation of what is causing the problem

https://rentry.org/hdgrecipes#bruising-in-strawberry-mix-pics
>>
>>7011800
tha'ts really hot, is there a special prompt for arms on chest like that? hands on own chest doesn't work that great for me
>>
>>7011791
>>7011750

I upscale 2x first using lollypop. This is very important because if you're inpainting small, high detail areas like eyes, the AI won't have room to do anything when the area you're inpainting is the size of an ant.

then only include the 'quality' tags like
>masterpiece, (highly detailed:1.2)

as well as a description of the area I'm trying to inpaint. Omit everything else except those very close to the area I'm inpainting, so context will be provided for the AI.

For example, if I'm masking just the eyes I would prompt these and these only:
>masterpiece, (highly detailed:1.2), face, (heart-shaped pupils:1.0), blonde hair,

Set denoising to 0.5-0.7 depending on how big of a change I'm making. If I'm just removing the visible seams after I've successfully inpainted what I want, I set denoise to 0.25-0.4. I keep CFG at 10, effects of changing it seem a little unpredictable.

The rest is
>DPM++ 2M Karras, Inpaint at full resolution, padding pixels 25-100, width and height at a constant 512x512.

Just typing all this shit out to save some time for people who've never inpainted and want to try it out. The above settings might not be the best, they're just what works for me at the moment.

Progression images for anyone who's curious how the whole thing goes: >https://www.pixiv.net/en/artworks/103084225

You can see some visible seams across images, those are the borders of the inpainting mask.
>>
cute
>>
>>7011822
the merged feet or the merged fingers?
>>
>>7011818
I'm actually using hands on own chest, but i always have to strengthen it to make it work.
>>
File: askdj324as.png (1.76 MB, 1024x1536)
1.76 MB
1.76 MB PNG
>>
>>
>>
>>7011828
Doesn't she get hot in that outfit? She must sweat a lot when she's outside...
>>
>>
>>7011838
>amputee
>>
did anyone do a bginga embedding or am I doing it
>>
A brief explination for what tags surrounded with () that end in :1.2 or :1.5 mean?
An alternative to saying (())? I'm unsure.
I am using NovelAI on browser; is this concept of parametre {breasts:1.2} compatible with what I have?
>>
>>7011847
() without number inside is a 1.1 multiplier. Default tag attention number for "tag" is 1.1 so (tag) will be 1.21, ((tag)) is 1.331. If you put number inside you can set attention lower or higher. "(tag:1.1)" is the same as just "tag".
>>
>>
>>7011847
>(n) = (n:1.1)
>((n)) = (n:1.21)
>(((n))) = (n:1.331)
>((((n)))) = (n:1.4641)
>(((((n)))) = (n:1.61051)
>((((((n)))))) = (n:1.771561)

>[n] = (n:0.9090909090909091)
>[[n]] = (n:0.8264462809917355)
>[[[n]]] = (n:0.7513148009015778)
>[[[[n]]]] = (n:0.6830134553650707)
>[[[[[n]]]]] = (n:0.6209213230591552)
>[[[[[[n]]]]]] = (n:0.5644739300537775)
>>
>>7011847
>I am using NovelAI on browser; is this concept of parametre {breasts:1.2} compatible with what I have?
No, on NovelAI you have to stick with using {{{n}}} in multiple layers.
>>
>>7011827
Loving the sketched look on this one, what was your prompt?
>>
>>7011847
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#attentionemphasis
>>
anyone try this yet?

https://github.com/Extraltodeus/multi-subject-render
>>
>>7011862
this looks extremely promising
>>
>>7011852
Disregard this, I suck cocks.
>>
>>7011867
Why do you do that?
>>
File: fasixasdas233.png (1.53 MB, 1024x1536)
1.53 MB
1.53 MB PNG
>>7011837
she doesn't mind working up a bit of a sweat.
>>
File: .jpg (149 KB, 1200x630)
149 KB
149 KB JPG
>>7011871
Cuz they're delicious?
>>
>>7011427
Elf knight-kun with her nervous "h-how do I look?" expression is so fucking adorable!
>>
>>7011862
kinda confusing to use
>>
>>7011877
Oh, my mistake. Please continue.
>>
File: .jpg (3 KB, 300x168)
3 KB
3 KB JPG
>>7011880
Did he feed he some spice or what?
>>
>>7011862
>https://github.com/Extraltodeus/multi-subject-render

Holy crap, gotta try this out. Thanks for bringing it up anon.
>>
>>7011862
This looks woefully insufficient. It generates each subject on a separate image and makes a collage, meaning we still can't have meaningful interaction (sex)
>>
>>7011880
And it still fucked up with the eyes the n my waifu
>>
>>7011891
Not exactly. It processes the multiple characters and background with a final prompt. In theory you could make them do whatever through that, just need to figure out how.

You can see in this image for example the girl is interacting with the guy >>7011880
>>
>>7011857
nsfw, bokeh, (highres:1.2), (ultra-detailed:1.2), (oil painting:0.9), (sketch:1.2), (extremely detailed CG unity 8k wallpaper:1.2), soft lighting, HDR, 1girl, facing viewer, thigh focus, (lying, on back:1.2), (spread legs:1.2), shiny skin, maid, (fairskin:1.3), purple hair, long hair, hair between eyes, bangs, detailed pupil, beautiful eyes, detailed eyes, expressive eyes, blue eyes, serious, (blush:0.8), perfect hands, huge breasts, puffy nipples, (breasts out:1.5), curvy, wide hips, maid headdress, black gloves, bowtie, black dress, short sleeves, apron, black thighhighs, garter straps, wedding ring, skindentation, no panties, (detailed pussy, pussy:1.5), (pussy juice:1.1), steaming body, (sweat), (((hands on own chest))), skirt lift, (((on a bed inside a luxury bedroom))), pillow, dynamic angle, looking at viewer,

Negative prompt: lowres, bad anatomy, bad hands, mutated hand, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (((idolmaster))), genshin impact, touhou, fate \(series\), colored pencil \(medium\), out of focus, glowing eyes, (wedding ring:1.1), (cleavage:1.5),
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 2673117376, Size: 960x1280, Model hash: 38c1ebe3, Model: Anything-V3.0-pruned-fp16, Variation seed: 1321736517, Variation seed strength: 0.05, Denoising strength: 0.65, Clip skip: 2, ENSD: 31337, First pass size: 448x640
>>
>>7011898
Right, but I think what it does is running a final img2img on the collage. In that case, it changed the cuff into the girl's hand. Getting any composition will be very challenging
>>
File: .jpg (671 KB, 896x1088)
671 KB
671 KB JPG
Have some 2b pixel art
>>
>>7011918
Yeah, but I feel like once someone figures out how to do one specific pose you could just swap out the characters.

I can also think of a bunch of uses that don't involve the characters from different prompts fucking. For example a threesome scene with one girl fucking and the other masturbating by herself.
>>
File: .jpg (729 KB, 896x1088)
729 KB
729 KB JPG
>>7011921
And watercolored 2b
>>
File: file.png (393 KB, 512x512)
393 KB
393 KB PNG
>>
File: image.jpg (839 KB, 2400x1800)
839 KB
839 KB JPG
Started off as a low effort inpaint of another artist's image that took too much effort to do. Results were pretty disappointing.

Original on left, inpainted on right.
>>
>>7011875
What are her hobbies anon?
>>
File: tmpy0k80tc2.png (577 KB, 704x1216)
577 KB
577 KB PNG
Why is a seiza so hard, it's always a tippy toe seiza instead of the regular one.
I was able to get it to work through photoshop but it really shouldn't be that hard... am I missing something?
>>
>>7011926
what prompts did you use for those splashes of watercolor
>>
>>7011946
(watercolor \(medium\):1.7)
you can play with weighting
>>
for embeddings, does anyone train on chinai fp32?

or do I just stick with NAI?

in both cases, I do not use the 7gb variant, does it matter alot?
>>
Thoughts about this?
>>
>>7011973
both nipples need inpainting to fix them, but otherwise not bad
>>
>>7011952
thx
>>
>>7011977
what prompt would i use?
>>
>>7011973
do people really like this over makeupped bimbo look? and her nipples are obliterating the hakama
>>
>>7011977
full prompt

ultra-detailed, masterpiece, absurdres, intricate, 8k, (high detail:1.1) , (painting by sophie anderson, bouguereau:1.2), saturated, wooden wall, wooden window, wooden floor,
Tifa lockhart, sweaty, open white (kimono:1.1), red (hakama:1.2), (long brown hair), hime cut, (1 girl with detailed face), lips, (large breasts:1.2), ( covered nipples:1.1), cleavage, smirk, (freckles:0.8), red lipstick, (make up)
Negative prompt: (((bad hands))), ((lowres)), (black and white), (((bad anatomy))), (((bad proportions))), (((multiple arms))), (((multiple legs))), (((multiple feet))), bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, blurry, username, out of focus, censorship, censored, ugly, old, deformed, amateur drawing, odd, fat, morphing, extra limbs, ((simple background)), (((depth of field))), extra limbs, (poorly drawn face), (colored sclera), (((multiple views))), monochrome
Steps: 24, CFG scale: 12, Seed: 3526719216, Size: 768x1280, Model hash: 925997e9, Denoising strength: 0.25, Clip skip: 2, Mask blur: 4, SD upscale overlap: 256, SD upscale upscaler: R-ESRGAN 4x+ Anime6B
>>
>>7011986
>inpainting
>>
>>7011991
you do need a prompt specifically for inpainting right? or should i just mask it with the full prompt
>>
>>7011987
mostly no, but sometimes seeing a non-slut character slathered in makeup like a whore is hot
>>
>>7011989
>>7011973
big fan of the freckles
>>
>>7011989
>>7011993
leave
>covered nipples
in positive prompt and put
>nipples
in negative prompt
then do this >>7011991
>>
>>7011987
>do people really like this over makeupped bimbo look?
yes
>>
Does anyone have any tips for getting unusual hair colors that don't have clear tags? Can you just color code it?

>>7011409
prompt?
>>
>>7012002
hex codes do not work
>>
>>
>>7012005
where's the banana
>>
>>7011998
Thanks (unrelated picrel i'm just preparing my next batch)
btw for inpainting, is it ok if I inpaint with the full prompt? I always thought I had to only prompt the specific area i'm inpainting
>>
>>7012002
That's not hard, mix two colors using : and | it takes a bit of fine tuning though.
The hard part is doing multicolored hair or highlights.
>>
File: you just know.png (1.15 MB, 1472x1472)
1.15 MB
1.15 MB PNG
>>
>>7012002
just photoshop it in, make a good image with just the main hair color, then use the paintbrush with a low hardness and an a medium to high opacity to color in the second hair color.
You can also use the paintbucket tool with opacity if you need to preserve the lines in the hair.
>>
>>7012018
im not learning how to use pshop. I asked cause I wanted an easy method, obviously im aware you can repaint shit using editing software.

>>7012015
not sure how these work, can you post a simple example?
>>
after playing around the various mixes and models, I think vanilla NAI is still the best for consistent sex prompts
>>
>>7012020
Then getting multi colored hair is a gamble. I was just giving the method that is sure to work.
>>
I'm training a hyper network right now to copy the style of a specific artist. Do I need to give my artist a name recognized as a word in the CLIP model or can it be anything?
>>
File: 112.jpg (2.95 MB, 2304x3072)
2.95 MB
2.95 MB JPG
>>
>>7012021
Strawberry is shit in general, doesn't really listen to prompts, looks good sometimes
berrymix actually listens, very good overall, closer to being an improvement to NAI but not always
nutmeg kinda like berry but a bit less reliable
As for porn I think NAI is about equal to berry and nut - they do produce prettier pictures though, just accuracy aint as good. Haven't tried the others.
>>
>>
those tactical thighs
>>
>>7012034
whats the prompt for this deliciously overweight elf whore?
>>
>>7012034
cute goblins. Prompt?

>>7012042
how'd u get a crowd this good?
>>
Is there any possibility that AI illustrations will be regulated in the future?
>>
File: 06394-SD_414077559.png (594 KB, 512x768)
594 KB
594 KB PNG
No nuns?
>>
>>7012055
>is there any possibility that something which has broad potential applications in the entertainment industry will be regulated
yes

>is there any possibility that someone will come to your house and rip the naked girls out of your computer
no ain't nobody got the money for that
>>
>still no model to make some Shizuka's
Truly the wrong timeline....
>>
>>7012055
Depends on the type of regulation you're referring to. Nobodies gonna take away your locally generated stuff but art sites might ban AI art and I wouldn't be surprised if future AI projects implement a subtle watermark to quickly identify AI art. The big projects will also have increasingly sanitized datasets to ensure no copyrighted, pornographic or offensive material is easily generated
>>
>>7012045
unfortunately complete rng from an ignored dynamic prompt and a custom model
>masterpiece, best quality, illustration, , 1girl, goblin, wetlook business suit, maid headdress, wariza, tank interior

>>7012046
same with this.. the crowd is from "busy restaurant in background" I guess
>>
>>7012009
Please someone answer, do i need to change prompts if i inpaint
>>
she looks so proud of herself
>>
>>7011987
>do people really like this over makeupped bimbo look?
Absolutely
>>
File: Murakamistyle.png (290 KB, 850x264)
290 KB
290 KB PNG
Night Shift nurses, the next generation. Accidentally put 200000 steps instead of 20000, but currently at 137/200000

This is the result after adding commas to my tags and manually editing the images, birme is kind of shit for cropping, i get better results resizing.
>>
based on my experience, it seems like embedding's strength over hypernetworks is stronger

what about you guys?
>>
>>7012072
>do i need to change prompts if i inpaint
depends on what you are trying to achieve:

>trying to get a new rng roll on an existing element (eg. redrawing a finger)
not really, keep the prompts as they are and pray for good rng.
>trying to remove something (eg. removing a nipple)
if you have prompts emphasizing given thing, you should remove them
>trying to add a new thing (eg. adding a nipple)
you can add it to the prompts and emphasize it.
I've also seen people remove the majority of their prompt and replacing it with the new thing, not really sure if one approach is better from the other.

denoising strength has big impact on inpaints (i usually inpaint at around 0.65 - 0.7).
if you are generating batches of inpaints, you should probably set your seed to -1
inpainting at full resolution is also helpful, i usually use it for finer details, but I've seen people always have it on (https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#inpaint-at-full-resolution)
>>
File: tmp7hah4wi0.png (1.08 MB, 960x1088)
1.08 MB
1.08 MB PNG
>>
>>7012084
What does embedding even do? When I open the train tab there is already a value in it, but i just usually delete it and leave it blank before setting hypernetwork to 0.00005.

Does it affect results?
>>
what's the difference between a wide shot and cowboy shot in tags?
>>
>>7011821
where can i find lollypop upscaler?
can someone help
>>
>>7012088
>Does it affect results?
training a hypernetwork? no. it is for training a whole separate thing

>>7012084
depends on the embedding, depends on the hypernetwork. can't generalize since they're all different. different datasets, different learning rates, different step count etc
>>
>>7012096
cowboy shot is when the girl is visible from the thighs and above and takes up most of the frame while a wide shot means any view where she doesn't take up most of the frame
>>
>>7012100
Good to know, thanks.
>>
>>7012100
post examples? trying to get my tag game up.
>>
>>7011589
Thanks for helping anon.
I wrote --precision gull on the web-ui-user.bat (on COMMANDLINE_ARGS), is that right?

The web ui opened with no problems:
"Launching Web UI with arguments: --precision full"

But when I tried to generate one image, i got en error.
This is the last two lines:
>attn_output = torch.bmm(attn_probs, value_states)
>RuntimeError: expected scalar type Half but found Float

Any tips on how solve it? Or if it is possible to solve it?
>>
>>7011217
prompt pls
>>
>>7012103
Add
>--no-half
along with
>--precision full
I want to say again that this will slow your generation down a noticable amount, although you'll still be generating on a 3060 compared to a 960.
>>
>>7012109
Right, it is like this
>set COMMANDLINE_ARGS=--precision full --no-half
The error is gone, but the picture stills like this >>7011502
Didn't got back to the old visual.

Non related pic
>>
>>7012114
>>
File: file.png (301 KB, 512x576)
301 KB
301 KB PNG
i like how this one turned out
>>
Anyone have the different sampling method and image output comparison screenshot?
>>
>>7012128
just make one yourself with the scripts...
>>
>>7012026
lovely style, could you share the prompt?
>>7012060
great sag
>>
What's this crap mean when training HN?
AssertionError: no gradient found for the trained weight after backward() for 10 steps in a row; this is a bug; training cannot continue
>>
>>7012088
embedding tries to find the hidden prompts inside the model that matches your training data
that's the simplest explanation
>>
the day a model consistently renders used condoms, is the day the machines take over
>>
>>7011973

Not bad! Glad to see someone else getting use out of my settings (after all, I got them from whom I will admiringly refer to as "Three Freckled Nun Blowjob" anon).

You could correct the issue with the nipples just by negative prompts.
>>
File: Clipboard5.jpg (633 KB, 2852x1225)
633 KB
633 KB JPG
>>7012129
I tried the x/y plot thing, the sampler doesn't seem to take effect. All 15 output are using DPM++ 2M sampler.
Am I doing something wrong?
>>
>>7012139
its weird that they work between different models better than anons will tell you they will
>most of anon's advice about them seems a bit useless, i've gotten better results at high step counts with a low training weight than changing training weights after x steps
>>
>>7012144
you aren't supposed to put spaces in the xy plot i don't think but i think there's some bug if you updated recently where it doesn't actually change samplers, might be able to pull or revert
>>
File: 00472-~1.png (1.66 MB, 1152x1536)
1.66 MB
1.66 MB PNG
>>7011879

Haha, I like to imagine she's got a something of a superiority complex, or is generally high strung, it's cute to picture. Pixiv seems to like it too, though not as much as her in the shower.

>"Yeah, I know they're big, asshole. I see them everyday. I'm asking you if they're distracting."
>>
>>7012144
try 10,11,12 instead of a range in case that's confusing it
>>
>>7012144
>Am I doing something wrong?
Sampler X/Y plots are currently busted, I had to roll back to a commit from November 12th to get it to work, 98947d1
>>
File: tmpd76e974s.png (1.82 MB, 1024x1536)
1.82 MB
1.82 MB PNG
>>
Embed bros, if I preprocess images using WebUI and the crops don't show the complete face of the character I'm training for, should I just crop manually?
>>
>>7012157 (me)
Also, what is an acceptable rate of loss? I have no idea what it means other than number go up bad.
>>
File: tmpiovrw4eh.png (446 KB, 512x768)
446 KB
446 KB PNG
3090 here. Any preferred commandline args? So far its just xformers, autolaunch, and deepdanbooru.
>Milfsuka for your time
>>
>>7012150
>Sampler X/Y plots are currently busted
found a post about it
https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/4860
>>
Guys, I can't stop generating pictures of Ereshkigal eating junk food. I think I've discovered a new fetish.
>>
>>7012162
I'm very mixed on xformers. I feel like that the optimization is still applying during training, even though the UI and such states that it is not, which leads to training frying very quickly.
It does speed things up, but I don't think it is necessary on a 3090.
>>
>>
>>
>>
>>
>>7012169
I ran a TI up to ~200k at .00005 without it breaking so I'm not sure what folks are too worried about unless things are different for hypernetworks (also you can turn things on and off when training much like nai's vae needing to be removed from models during)
>>
>>
File: 1661785985381.png (547 KB, 512x768)
547 KB
547 KB PNG
>>7011821
Not any of those anons but I appreciate your reply to him. I used to attempt to do heart shaped pupils but it always came out bad so I stopped. Looks like I have to learn to inpaint. If only there were full prompts on the progression images for the inpainting steps and how the masks should look like but that's already solid enough advice. Thank you.
Also I checked out your stuff, I really like the multiple character images. How were you able to achieve those results? Usually my multiple character images look terrible and with that high res the details are ven worse. I really like 103063882. And then there's 103064714 has all of them touching each other without melting into each other. Did you inpaint those as well perhaps?
>>
File: Definitive.png (849 KB, 1024x1024)
849 KB
849 KB PNG
>>
>>7012175
I think it comes down to that things really aren't implemented properly in the WebUI at the moment. TI Training and Hypernetwork training are both implemented wrong on the main build. The VAE selector doesn't actually fully unload the VAE, and you can tell that the minute that you start training.
I wish the updates were more frequent, like there were a few weeks ago, but I can understand that the guy is probably feeling burn out.
>>
File: Finished.jpg (1.05 MB, 3072x5632)
1.05 MB
1.05 MB JPG
Finished a new work in my impossible series, a bit dissatisfied with this one, but it's okay.
>>
File: 00587-~1.png (1.35 MB, 768x1600)
1.35 MB
1.35 MB PNG
>>7012184

Nice. Vaguely Studio Trigger esque.
>>
>>7012182
Nope, I only inpainted small details like hands, eyes and nipples. If I had to inpaint that large of an area I would just generate a new picture.

If you PNGinfo the image from pixiv, you'll be able to see the whole prompt and generate your own. Just don't stare too long at the negative prompts.

Model is the unpruned AnythingV3 model.
>>
>>7012188
>TI Training and Hypernetwork training are both implemented wrong on the main build
how so?
>>
File: 00024-3434742484.png (1.64 MB, 1024x1536)
1.64 MB
1.64 MB PNG
>>
>>7012188
I've done things with the webui's TI that I wouldn't be able to do if it wasn't working (pic related); what exactly is supposed to be wrong with it?
>>
anyone got a good hololive wildcard? the one in sdupdates is shite for half of them
>>
>>7012191
Love her! What was this prompt?

I tried to recreate Elf-kun with the prompt you posted yesterday or the day before, but she came out a bit feep fried. I might try again later and if it's still messed up with this prompt, I probably have some settings that are off.
>>
>>7012193
>>7012196
I'm basing that off a SD Update from a few days ago
https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/4680
What currently is implemented does work, but I feel like switching to this build, or using that new Hypernetwork extension, has led to better results for me. It's all personal opinion.
>>
>>7012202
"there is a new better option" doesn't mean "wrong," it means "out of date"
>>
Read through the WebUI README for the first time, didnt know about highlighting a word and typing ctrl+up and ctrl+down...

>>7012202
>Hypernetwork extension
What's it called and why is it better than the current WebUI implementation?
>>
>>7012196
It's more like half of the functions are broken or are just not there. Dropout and sampling especially. This discussion goes over some of the things wrong right now:

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/4886

After adding things from there my hypernetworks have become a lot more consistent in their trainong
>>
>>
>>
>>
File: 00580-~1.png (1.24 MB, 960x1280)
1.24 MB
1.24 MB PNG
>>7012200

Okay, I'd like to see it if you still have it, lol. I think the first appearance of Blonde Elf Knight-kun on the booru was actually someone who did their own flavored spin-off with intricate skin patterns/transparent sleeves, before any of mine were uploaded. I suppose it depends on what very subtly different flavor of Saori you were trying to cook. Even on my Pixiv I think she has the most entries now, outnumbering Arshes Nei.

Anyway, here's workout Blonde Elf Knight-kun. She's not even wearing an aggressive tight workout thong, she practically looks like your average muscular blue-eyed blonde with abs and fat tits. Cheers!

4k, highres, (masterpiece:1.2), (best quality:1.2), ultra realistic, (detailed:1.2), (feminine:1.3), sexy elf woman, pointy ears, blonde hair, (short braids:1.2), bangs, (bright blue eyes:1.1), frown, large breasts, (cleavage:1.1), (mature female:1.1),round hips, narrow waist, skin indentation, (breast conscious:1.2), (covered nipples:0.9), (1girl:1.1), (sexy workout clothes:1.2), jewelry, (abs:0.8), (athletic:1.4)
Negative prompt: lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,deformed face, deformed eyes, poorly drawn face, (ponytail hairstyle),((bad anatomy)), ((bad proportions)), ((multiple legs)), ((multiple arms)), out of frame,(((monochrome)))
Steps: 28, Sampler: Euler a, CFG scale: 11, Seed: 588937006, Size: 768x1600, Model hash: 925997e9, Denoising strength: 0.7, First pass size: 448x640
>>
>>
>>7011811
>Is there a way I can inspect the images here?
As another anon said 4chins strips metadata so you can't get it from random shared images -- unless you're on a photography board.
If they have added a catbox or imgur link, then yes. Which IMO is much cleaner than copy-pasting a wall of prompt text

Maybe it's time to nag moot2 for a green /ai/ board? Though the /b/ thread would still have to exist for loli posters
>>
>>7012216
An option to allow keep metadata in the options field would be great
>>
>>7012224
>>7012216
Honestly a new thread or a keep meta option would both be amazing.
>>
>>
>>7012211
Here's what came out of it when I tried it lol. Not the intention, I reckon.

prompt: (masterpiece:1.3), (best quality:1.1), (ultra realistic:1.2), (detailed:1.2), (feminine:1.3), pointy ears, (blonde hair:1.1),(short braids:1.2), (bright blue eyes:1.1), smirk, large breasts, cleavage, mature female, round hips, (narrow waist:1.2), skin indentation, (bangs:1.2), (covered nipples:0.9), (1girl:1.1),solo focus, young elf schoolgirl, jewelry, blue microbikini, muscular, abs, (athletic:1.1), hallway background
Negative prompt: lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,deformed face, deformed eyes, poorly drawn face, (ponytail hairstyle),((bad anatomy)), ((bad proportions)), ((multiple legs)), ((multiple arms)), out of frame,(((monochrome)))
Steps: 28, Sampler: Euler a, CFG scale: 11, Seed: 2784641961, Size: 448x640, Model hash: 925997e9, Hypernet: cutesexyrobutts, Hypernet strength: 0.25, Clip skip: 2, ENSD: 31337

I'll try again with your new prompt
>>
>>
Can people post a sample prompt of some sex position?
I want to see where you guys are positioning your prompts so it remembers there's supposed to be a guy in the pic.
>>
>>7012246
1boy lmao
>>
File: 00005-~2.png (1.34 MB, 768x1600)
1.34 MB
1.34 MB PNG
>>7012234

Oh god, that trail from her navel to her crotch. I'd be horrified if I wasn't laughing so much.

I think...that might just be a bad seed. And in need of a hypernetwork. She vaguely resembles some of the more "realistic but shiny" outputs I get, like some sort of weird Bodybuilding Contest version of Saori that I only keep as an experimental basis towards less stylistic versions.
>>
File: 00001-588937006-2.png (1.54 MB, 1536x1024)
1.54 MB
1.54 MB PNG
Actually, looking at the settings data--it must entirely be that seed combined with that particular hypernet (which isn't a bad one, I've used it for other things, just not any of my mainstay Elf Pin-ups). My go-to seed is surprisingly resilient to hypernetwork influence, so it really must be that one. Otherwise everything looks correct, right down to that obsolete (ponytail hairstyle) negative prompt.
>>
First class service for you, sir?
>>7012246
It depends on the seed in my experience.
>>7012248
Doesn't always work, especially if you've detailed the woman a lot. Can end up making the girl boyish or generating some lesbomancy.
>>
>shitload of errors when running interrogation
>install all requested libraries
>interrogation no longer functions
idgi
>>
>>7012194
Hot. Prompt?
>>
>>7012250
Yeah, I thought it might be that.

In other news, even with inpainting, it continues to be extremely difficult to get good feet. The best alternative I've found so far, is to generate an image that's about what you'd like, photoshop in a foot from an artist who can actually draw them, and then inpaint on top of it with low denoising strength. Socks tend to make it easier.
>>
>>7012269
is that nai?
also prompt pls
>>
>>
>>7012152
looks really nice, good job anon
>>
>>7012269
Yep, NAI! the prompt works really well to generate one specific pose, but yeah even with it the feet are just impossible to get right. Image here is what I got before 'shopping it.

nsfw, bokeh, (highres:1.2), (sketch:1.2), (extremely detailed CG unity 8k wallpaper:1.2), HDR, 1girl, facing viewer, foot focus, (lying, on back:1.2), (spread legs:1.2), heavy makeup, (gyaru makeup:1.3), see-through tank top, (dark skin:1.4), long blonde hair, drill hair, hime bangs, detailed pupil, beautiful eyes, detailed eyes, expressive eyes, pink eyes, elf, pointy ears, smug smile, (blush), perfect hands, perfect feet, cleavage, covered nipples, ((curvy)), wide hips, skindentation, (wet pussy), pubic hair, ((steaming body)), (steam), ((sweat)), (((arms up))), (((legs up))), (on a bed in a bedroom), pillow, dynamic angle, looking at viewer, ((full body shot)), (((soles, white socks)))
Negative prompt: lowres, bad anatomy, bad hands, mutated hand, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, out of focus, glowing eyes, ((feet out of frame)), (deformed feet), (mutated feet), uncovered nipples, more than 2 nipples, shoes, panties
Steps: 28, Sampler: DPM++ 2M Karras, CFG scale: 16, Seed: 4226571762, Size: 448x640, Model hash: 925997e9, Hypernet: anime, Hypernet strength: 0.25, Batch size: 8, Batch pos: 6, Clip skip: 2, ENSD: 31337
>>
Prompt ideas?
>>
>>7012143
I've been playing with your prooompt for a while now
please enlighten me with more senpaiii
>>
>>7012316
What model is that?
>>
>>7012322
This is just NAI
this is the prooompt
>>7011989
>>
File: 1637968699051.png (496 KB, 512x768)
496 KB
496 KB PNG
>>7012192
Oh okay, so the characters don't need to be inpainted then. That's exciting! Got the prompt and the hypernet too but I wasn't able to get the generation though. It seems like some of the settings are incomplete in the data.

For instance, 103063882 doesn't have the sampler used, though I assume that's DPM++ 2M Karras as per your posts. The size is also 512x512 which is incorrect but seems to coincide with your inpainting. The positive text content also seems to be too short, and seems to be an inpainting style prompt. What I'm guessing with most certainity is that it grabbed the prompt for your last inpainting attempt and overwrote the original prompt. All of the images seems to be like that. Do you mind posting the original prompts for them if you can? Thanks again.
>>
File: Clipboard7.jpg (1.74 MB, 4000x2311)
1.74 MB
1.74 MB JPG
>>
File: darj paizuri msdos.gif (1.55 MB, 512x512)
1.55 MB
1.55 MB GIF
I reused a few outputs from testing to make a .gif
looks okay for being made from scraps. i added a filter to hide imperfections and make the banding from the .gif palette conversion less out of place. should've used dithering instead, but didn't.
>>
>>7012329
https://files.catbox.moe/ml35nb.png
https://files.catbox.moe/xakpfu.png

Here's what I could find, I think I deleted the rest of the originals after I was done inpainting them.
>>
>>7012131
proooompt please?
>>
>>7012316

The sagely Elf-Anon sezs: I don't think it's as good for mikos as it is for...western girls? girls in blazers?, because it's very good at drawing a bra peeking out of a packed blouse, and you don't generally get that with kimonos, at least not as easily.

Also, she looks like she spilled sugar syrup on herself rather than cum, but that's not necessarily a bad thing (it'll taste better for all involved, but you do risk getting ants). Due our inevitably exaggerated expectations of "cum on face."
>>
What are you staring at, human fool?
>>
But you're doing well with see-through tops. I cut my own efforts short and have been meaning to return to them.
>>
File: Murakamistyle.png (686 KB, 1705x921)
686 KB
686 KB PNG
Okay, so after spending 3 days tagging, fixing tags, finding clips, cropping etc. I'v finally completed my dataset.

Before i begin I have an RTX 6600 with 8GB Vram, please recommend me the best settings to train this thing so I can finally take a break.

>learning rate
>steps
># of results per steps

I'm currently using Anything V3 pruned, is there a better model to train it?
>>
File: tmp2em_kw74.png (674 KB, 640x960)
674 KB
674 KB PNG
Figuring out what those damn little shoes were called was a godsend.
uwabaki btw
>>
>>7012343
prooompt for this outfit?
>>
>>7012346
First you have to say what the fuck you're training.
>>
>>7012346
>RTX 6600 with 8GB Vram
You are the first anon in these threads that I know of who is trying to train on AMD hardware. I think this is the kind of question you have to ask in the WebUI github repo
>>
File: murak-3600.png (54 KB, 911x545)
54 KB
54 KB PNG
Are these settings good? Should I select anything?
>>
>>7012350
bro it's his namefag, the name of the fucking image, and you can see the file names. are you retarded anon?
>>
File: 00948-~1.png (1.27 MB, 1024x1024)
1.27 MB
1.27 MB PNG
>>7012344

That's a tough balance between "sexy orc girl" and "statuesque vegetable woman", as countless Skyrim mods demonstrated. I think this is a good example of the later, though you could probably get away with making her hair and genitals not green if you so desired.

Also, I like the gloves and possible war hammer just out of frame, even if it almost looks lik she has six fingers.

>She's very happy to have her spear.
>>
>>7012350
I'm training a hypernetwork on the artstyle of the hentai director Murakami Teruaki. I have about 350 images, all cropped and tagged by hand.

https://www.pixiv.net/en/artworks/45610059

These are results I'v gotten over the past few days during test runs.
>>
>>7012131
prompt also please
>>
Absolutely disgusting
>>
>>7012353
No shit. I'm talking about what he's training not the subject. Dreambooth, embed, hypernet.
>>
>>7012354
Yeah, the hair is supposed to be black, but the green skin prompt just bleeds into everything
>>
>>7012356
these webms/gifs are all ai generated?
>>
This is a bit better, I think. After all, orcs are mushrooms, not vegetables
>>
>>7012361
No these are taken from various series he's directed. Im training the AI on his style so i can make my own stuff.
>>
Do people mostly use DPM++ 2M Karras? I just keep switching between Euler and Euler a...
>>
>>7012346
>anything v3
look I like the chinkai as much as the next anon but trying to bend anything will only amount to despair. I think you should use NAI instead
>>
>>7012365
is there a renpy or website that explains the different sampling methods? i used to be able to mouse over them in automatic1111 and it doesnt give that mouse over hints anymore
>>
>>7012360

Whoops, I meant to say "example of the former"; but yeah, she kind of looks like a giant vegetable as much as a sexy orc lady.

You could try (green hair)--or for that matter (green nipples) in the negative prompts of varying strength, if you haven't.

>>7012349

Well, because you asked so nicely (I can't take all the credit anyway). It's really easy to mess up her arms, as you can see.

ultra-detailed, masterpiece, absurdres, intricate, 8k, (high detail:1.1) , (painting by sophie anderson, bouguereau:1.2), saturated, wooden wall, wooden window, wooden floor,
sexy schoolgirl, red (school uniform), ((school blazer)), (long blonde hair), hime cut, (1 girl with detailed face), (pointed ears:1.1), lips, (large breasts:1.1), cleavage, bra peek, smirk, (freckles:0.8), red lipstick, (make up)
Negative prompt: (((bad hands))), ((lowres)), (black and white), (((bad anatomy))), (((bad proportions))), (((multiple arms))), (((multiple legs))), (((multiple feet))), bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, blurry, username, out of focus, censorship, censored, ugly, old, deformed, amateur drawing, odd, fat, morphing, extra limbs, ((simple background)), (((depth of field))), extra limbs, (poorly drawn face), (colored sclera), (((multiple views))), monochrome
Steps: 24, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 434227704, Size: 768x1280, Model hash: 925997e9, Denoising strength: 0.6, Clip skip: 2, First pass size: 0x0

Though I like how you've gone from "busty miko" to "rich big-boobed sempai in a hakama wearing a backpack."
>>
>>7012356
...why did you randomly link to a pixiv animation from 2014?
>>
>>7012368
I'd second this. NAI is a good baseline to train on, since the hypernetwork trained on NAI will play well with Anything or other models due to NAI being the baseline.
It's going to be real fun come December when everyone might have to start over if WD1.4 ends up being as good as they're hyping up.
>>
>>7012365
we use dpm++ 2m karras because it can produce great images in only 20 steps (15 or sometimes even 10) which means you can cut production time way down and allow you a ton of gens quickly without any noticeable lost of quality
>>
>>7012368
Can you drop a link to the repo? Iv tried googling but not sure any links are right
>>
>>7012356
>>7012361
>>7012372

Sorry, I'm retarded.

https://www.pixiv.net/en/users/9936494
>>
>>7012365
i use non karras
>>
>>7012371
Yeah, I just threw a bunch of brackets around (((black hair))) actually
>>
>>7012371
I'm getting a good insight about how you're engineering this, thank you kind sir.
>>
my 1080ti has 11gb of vram, but depending on settings and how long i've been running it. it runs out of vram. whats the best way to clear it out without rebooting my entire rig. albeit a reboot is only like a couple seconds
>>
>>7012382
>>
>>7012384
thank you based chad
>>
File: unknown.jpg (1.91 MB, 4224x1676)
1.91 MB
1.91 MB JPG
>>7012374
>he isn't using the superior SDE Karras at 5 steps
lol
>>
>>7012384
looks like a reboot would still be quicker KEKW
>>
>>7012381

Cheers! The common problem of the AI failing to distinguish between layers of clothing is still...common (artists too, I suppose); so three layers (blazer, blouse, bra) can easily become two. I haven't gotten around to putting any of my miko images of any style in my Pixiv yet, though I'll probably do a wide review of that when I do. Still on swimsuits and catsuits.
>>
File: cS2fGi2.jpg (31 KB, 375x199)
31 KB
31 KB JPG
>>7012373
Which one?
>>
File: tmpadsrtqi5.png (1.25 MB, 640x1024)
1.25 MB
1.25 MB PNG
What exactly is a hypernetwork and how can it improve my stuff?
>>
>>7012386
didn't know SDE Karras existed till this point, so that doesn't appear on my list and I'm up to date so I'm guessing you added it yourself? any link?
>>
For a couple weeks now, I've been running x/y prompts to find good combos of quality tags, with alternating samplers, various sampling steps and CFG, even S/R of said quality tags..
I've tried a bunch of models and merges..
But as time passes, I feel as if I've been getting worse at text2img instead of better. Pic here is probably my best result ever, 2.5 weeks ago. For the sake of clarity:

Positive: Tohsaka Rin standing at the beach wearing black and red (slingshot swimsuit), hand on crotch, grabbing own breast, NSFW, 1girl, solo, ((full body)), long legs, halterneck, (sling bikini top), aqua eyes, large breasts, clenched teeth, blushing, masterpiece, best quality, ultra realistic, detailed, 4k, 8k, cinematic, official art, high resolution, high resolution illustration, extremely detailed, intricate details, lush detail, insanely detailed, hyper detailed. highres, 5 fingers, photorealistic, cinematic lighting, cinematic angle, vivid colors, amazing composition, trending on artstation HQ
Negative prompt is a full 18 lines of some crap I copied from somewhere, you really don't want to see what's in there..

Steps: 50, Sampler: DPM++ 2M, CFG scale: 16, Seed: 2469947653, Face restoration: GFPGAN, Size: 768x1024, Model hash: 35279b19, Model: BerryMix, Denoising strength: 0.5, Eta: 0.5, Clip skip: 2, First pass size: 512x768

I've only been making Rin images since I started. I figured since I'm trying to figure things out, at least I can spam one of my faves.

What are the most widely acceptable "default" good quality and bad quality prompts that just work with everything? Am I fucking myself over by maniacally trying to jam in as many quality tags upfront? Does this kill my chances to attempt something with a more "drawn" feeling?
>>
>>7012386
how do i get this sampling method?
>>
>>7012394
There was mention of it in /g.

https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/4961

I don't have it either, only now just googled it, so I can't attest to anything it claims or offer installation instructions.
>>
>>7012389
You might need to download it. The hash you're looking for is 925997e9, which I don't see in your list.
>>
>>7012340
>>7012357
makes me happy to see that someone actually likes the image. I've begun to question if people just don't like this style.

anyway, the prompt is:
ultra-detailed, masterpiece, absurdres, intricate, 8k, (high detail:1.1), detailed background, (painting by Hendrik Willem Mesdag:1.2), saturated, (realistic:1.1), fantasy,
nfsw, (((vaginal))), (penis), spread legs, 1boy, riding, pussy, ((pov crotch)), cum, pussy juice, ((vaginal penetration)), ((nipple slip)), (((girl on top))), ((straddling)),
1 girl, elf with perfect face, large boobs, (quin tails), purple hair,
pantyhose, unbuttoned shirt, corset, cleavage, see-through robes, (black lace-up boots), witch hat, wet shirt,
smug grin, freckles, (crazy eyes),
looking at viewer, looking down, from below, legs spread, torn legwear, squatting,
park in a medieval city, tavern, (((night))), wooden benches, stone path, bushes, flowers, trees,
Negative prompt: ((lowres)), (black and white), (((bad anatomy))), (((bad proportions))), (((bad hands))), (((multiple arms))), (((multiple legs))), (((multiple feet))), bar censor, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, blurry, username, out of focus, censorship, censored, ugly, old, deformed, amateur drawing, odd, fat, morphing, extra limbs, ((simple background)), (((depth of field))), extra limbs, (poorly drawn face), (colored sclera), (((multiple views))), monochrome, (((sad)), ((crying)), (((shy))), (((standing))), child, teen, small breasts, legs together, ((day)), daylight, from below, deformed penis, floating penis, detached penis, long penis, cow print, cow ears, cow horns, arms up, hands behind back, arms behind head
Steps: 24, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 2121445172, Size: 640x1152, Model hash: 925997e9, Hypernet: cutesexyrobutts, Hypernet strength: 0.5, Denoising strength: 0.7, ENSD: 31337, First pass size: 0x0
>>
>>7012396
yeah I found this page but I'm not exactly sure how to change branches on git to get this to pull
>>
anyone got a backup link for part 1 of the novel ai leak?
>>
>>7012398
i always feel greedy for asking for prompts and then get bummed when OPs aren't in the threads anymore or ignore the request so appreciate the prompt. its a fucking beautiful image. she's super sexy
>>
File: 1.png (14 KB, 913x180)
14 KB
14 KB PNG
Discussion on the sampler seems to indicate that it's not necessarily faster due to calling the model twice for each step.

Is there a noticeable quality increase in the same amount of time taken to generate something?
>>
yo chan chads.

is there a website or renpy that explains the sampling steps that SD has? i'm trying to find one in google but it keeps trying to send me youtube videos and i just want a reference page
>>
>>7012393
Sample of today's trials, for comparison and proof of quality drop.

Pos.: illustration of (tohsaka rin) at the beach with silver slingshot swimsuit, with white sun hat, barefoot. 1girl, (teenage), young girl, (full body). medium breasts, cleavage, cleavage cutout, sideboob, navel cutout, thick thighs, long legs. masterpiece, best quality, ultra realistic, detailed, 4k, 8k, cinematic, official art, high resolution, high resolution illustration, extremely detailed, intricate details, lush detail, insanely detailed, hyper detailed. highres, 5 fingers, photorealistic, cinematic lighting, cinematic angle, vivid colors, amazing composition, trending on artstation HQ
Neg.: multiple views, multiple girls, 2girls, 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, out of focus, censorship, ugly, old, deformed, drawing, odd, fat, adult, anime, cell shading, black and white, red eyes, huge breasts, large breasts, white hair
Steps: 40, Sampler: DPM++ 2M, CFG scale: 11, Seed: 3167822510, Face restoration: CodeFormer, Size: 512x768, Model: Nutmeg mix, Batch size: 2, Batch pos: 1, Eta: 0.2, Clip skip: 2, ENSD: 420619

Feel free to call me an idiot, btw, if you can add anything constructive at all, I don't mind. It takes skill to get WORSE with practice.
>>
File: file.png (551 KB, 512x832)
551 KB
551 KB PNG
kicking tags surprisingly work, high kick, flying kick.
Just make sure to put feet/shoes in the negative and standing on one leg in positives
>>
>>7012397
where????
>>
File: 1643869870061.png (111 KB, 1245x877)
111 KB
111 KB PNG
>>7012339
No problem! Thank you. Tried the 1st image and somehow missing the entire third girl and a bunch of the cute background detail already. All the settings seem to be correct and the hypernetwork is being applied. I've got all the standard settings like Clip Skip 2, eta noise seed delta 31337, and so on. I'm on the latest commit too. The prompt output is exactly the same as the reference too, as per below,

https://files.catbox.moe/c7eac0.png

I wonder if there are some hidden settings in the options that are not being reflected. Tried with and without vae to no effect. With and without quantization. With and without xformers. no real difference. The hypernetwork version could be different but its just the one I picked up from that one MEGA korean hypernetwork folder everyone uses and I presume its also the exact one you used. Sorry to bother you so much but do you mind checking what commit you are on and or the settings you use?
>>
>>
>>7012399
1. Go to the tab "Files changed" at the link provided.
2. Go to your stable diffusion webui folder in your PC
3. Find the 2 .py and the 2 .txt files mentioned in the "Files changed" tab.
4. Open each one with a program that allows you to edit them, something like Notebook
5. Make the changes (in "launch.py" you have to change line 148, in the other three you have to add an extra line) as shown in the "Files changed" tab. Be sure to pay attention to the line numbering, your Notebook should probably mention the line you are in at the bottom right.

Either take note so you can change everything back the way it was, or copy the four files and save them somewhere else before you start editing. You may need to revert the change at some point.
>>
>>7012387
this isn't twitch
>>
>>7012413
WeirdChamp
>>
>>7012405
Not sure if it'll be an improvement, but i've tried reformatting your prompt (outputting other stuff right now so no image). I like to keep prompts smaller (sub 75 tokens) so this should give you flexibility to alter details or adjust the composition/angle. Your negative is fine and can be fine-tuned depending on the seed you find (or you could try a schizo negative)

hyperrealistic, ultra-detailed, absurdres, (perfect face:1.1), intricate (high detail:1.1) body, cinematic lighting
1girl, (tohsaka rin), (full body), beach
(teenage),medium breasts, cleavage, sideboob, thick thighs, long legs, barefoot
silver slingshot swimsuit, white sun hat
>>
>>7012412
I'm getting a no module error now for torchsde, I did a pip install torchsde but that didn't seem to work
>>
File: 2.png (82 KB, 829x869)
82 KB
82 KB PNG
>>7012410
>MEGA korean hypernetwork folder
That's the one,yeah.

Using quantization, xformers and NAI's vae.

Currently on 828438b4a190759807f9054932cae3a8b880ddf1.

>https://files.catbox.moe/c7eac0.png
>Upscale latent space image when doing hires. fix

Did you enable that? It looks like you didn't.
>>
>>7012425
if you checked even the first couple of links in the OP
>>
>>7012427
turn off img2img color correction it will fuck your shit up
>>
>>7012412
If i'm not mistaken, to get locally the changes from a PR you should
- git pull (on your current branch)
- git stash --include-untracked (if you've made any changes on the branch and want to go back to them later, otherwise could be ignored)
- git fetch origin pull/4691/head:DPM++SDE
- git checkout DPM++SDE

and you are done. If you want to go back do.
- git checkout master
- git stash apply (if you want to get your previous changes back)
>>
>>7012429
Thanks anon, will do. What does it fuck up exactly? I've not done a comparison as I rarely use img2img, only do inpainting.
>>
>>7012433
do you have any tips? the process seems dum and gay a lot of the time

outpainting would be good too as that makes even less sense
>>
>>7012433
i'm honestly not too sure except it makes the colors different when img2imging...obviously it's supposed to be a positive feature, but i liked what i got better after i turned it off, and to boot the colors looked different when using it. since highresfix is a type of img2img it affects that too
>>
>>7012431
this worked, thank you anon,
>>
Is there a way to make it keep generating while I'm on another tab? it's kind of driving me insand having to just stare at the fucking progress bar every single time.
>>
>>7012438
anon its not based on the browser, its in the command prompt window...
>>
>>7012428
tried the magnet links, they're dead
>>
>>7012438
it should do that by default, if it isn't you've done something very wrong
>>
>>7012440
It's literally not generating unless I'm on the browser tab.
>>
>>7012416
> sub 75 tokens
Unless I cut majority of the supposed "quality tags", those alone are probably more than 75... Should I be lurking here daily? Did I miss the meta change? Have we stopped adding things like "masterpiece", trending on artstation etc?

I have zero qualms about using the schizo negative. I even got the "bad prompt" embedding. But doesn't overloading the negative also affect the resulting pic in a bad way?

P.S. I know I should do it "the proper way" i.e. batch in txt2img then choose one for img2img, then photoshop or equivalent, then inpaint, but my base txt2img really bums me out.

One more Rin for the road.
>>
>>
>>7012445
you're telling me the loading bar in the command prompt literally wont move unless you also have the browser tabbed in? then reinstall windows anon your shit is broke
>>
A different kind of orc massage
>>
>>7012435
See my reply in >>7011821.

It's a dum and gay iterative process, you will want to send to inpaint the moment you notice the smallest improvement that you want. You won't get what you want in <5 steps unless you get lucky with the seed/already made a decent drawing to inpaint over.

When drawing shitty mspaint objects, also use clear lines to divide objects, just don't draw them too thick.
>>
>>7012453
I never considered upscaling before inpainting, I'll have to try this on my next inpaint job
>>
File: tmpj9_ex1pb3.png (722 KB, 425x768)
722 KB
722 KB PNG
>>
it's been years since I've generated some Rin...
>>
>>7011207
Damn that looks amazing, what Mix did you use?
>>
>>7012446
There is no real unified "meta" at the moment. Some people swear by more than 75 tokens, others use masterpiece, some use sentence based prompting. The structure I use is:

Quality Prompts (sometimes styles/masterpiece/etc is added)
Image composition (Number of people, angle, location, etc)
Physical characteristics (hair, bust, facial expression, etc)
Clothing

Its not perfect by any stretch, but it works 90% of the time for me. Main modification is adding something higher up if its not getting noticed (or just adjusting the emphasis)
>>
>>7012442
Not dead, tested it right now. It's working, but your torrent client might have bugged out and gotten stuck in "downloading metadata." For some reason it has happened to a lot of anons.

Try this, it's the actual torrent file
https://files.catbox.moe/tk686z.torrent
>>
>>7012445
thank you anon!
>>
>>7012390
For all I know it's a file filled with dark magic that lets you either add a character that the model wasn't trained on or mimic the art style of a particular artist.
>>
File: 1655991662796.png (73 KB, 625x808)
73 KB
73 KB PNG
>>7012427
Looks like the same commit. Tried using NAI's vae this time, no difference. they appear to be same file. Slapped on quantization and xformers at the same time, doesn't seem to change anything.

>Upscale latent space image when doing hires. fix
>Did you enable that? It looks like you didn't.
Oh you are right! I enabled that and I got better results,

https://files.catbox.moe/bcp70d.png

However the third girl is still MIA. Where could she have gone!? These are my settings for the upscaled options. I wasn't even aware they would affect highres fix. Thank you for letting me know about that latent space image fix.
>>
>>7012470
woops, meant for >>7012401
>>
>>7012475
do you got any more of your sexy girl or other prompts you used on her?
>>
File: overlap.jpg (83 KB, 589x813)
83 KB
83 KB JPG
>>7012446
see >>7010773 and this image

the general order that works reliably is
styles
camera
pose
description
background/lighting

but you can sprinkle things through and prompt some odd things like close-up at the start of >>7012209 / https://files.catbox.moe/81oxlx.png and it will work fine
>>
>>7012474
What card are you on? I'm on a 3070.
>>
>>
>>7012410
>>7012427
>MEGA korean hypernetwork
What the hell is that?
>>
>>7012402
testing now...
I think the X/Y Plot doesn't see the SDE karras correctly because my times when generating were identical which didn't make sense, also the images were a little too close, so now I'm doing some individual testing
5 step:
DPM++ 2M Karras: 9 seconds
DPM++ SDE Karras: 9 seconds
10 step:
DPM++ 2M Karras: 10 seconds
DPM++ SDE Karras: 17 seconds
20 step:
DPM++ 2M Karras: 19 seconds
DPM++ SDE Karras: 34 seconds

so not faster, I'm still running tests at 5 steps to confirm some things but the jury is still out
>>
File: 1650627717133.png (57 KB, 457x669)
57 KB
57 KB PNG
>>7012480
RTX A6000. I don't think it would make that huge a difference, the difference level would be like using or not using xformers. I say that since I used different cards before and the results are usually near identical even when using xformers which always results in a slightly different image. I can however get a hold of an 3070 today and test it if you are curious. I'm not sure what else I can try to mess around with.

Fiddled with the other settings in upscaling such as using different img2img upscalers just in case it was somehow silently applying those to the high res fix option. But so far no notable difference aside from the upscale latent space fix for high res.

I don't have restore faces option on either, and I tried using both codeformers and GFPGAN but it doesn't do anything.
>>7012485
/folder/sSACBAgC#kNiPVzRwnuzs8JClovS1Tw/folder/dSQiRZQT
>>
>>7012477
I've a lot of images generated (from that prompt + wildcards) but I'm yet to look through them and pick out the good ones.
I also try to avoid posting ones with obviously bad anatomy without inpainting them first but since you are asking, i will make an exception.
I've posted in previous threads too, but it's usually one or two images per thread since inpainting takes a lot of time.
>>
>>7012495
An anon did some tests in /g/ and it doesn't seem as accurate generally to what he was used to with 2m karras, but on the plus side it had coherent results at only 4 steps or so, and it also made lines which the ai traditionally drew kind of squiggly, such as skyscrapers, nice and straight
>>
>>7012498
what are these wildcards people keep speaking of?

i only have positive and negative prompt sections in my automatic1111

is this a section i can enable?
>>
>>7012500

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Extensions#wildcards
>>
>>7012500
https://github.com/AUTOMATIC1111/stable-diffusion-webui-wildcards/
>>
>>7012503
>>7012504
thank you based chads. the /b/ thread moves too fast when i ask them questions it gets overlook a lot
>>
>>7012499
https://files.catbox.moe/bq7nee.jpg
5 steps at 7 CFG showed some pretty great results, left is SDE, right is 2M, I found at 11 CFG both were unusable, I did batch size 3 and found 2M was faster, but most of its gen was unusable.

not bad for 5 steps imo
>>
>>7012496
Yeah, I'm beat on what could possibly be causing the whole missing girl.

>I don't think it would make that huge a difference
I concur, sometimes the small differences are enough to make physical deformities appear compared to original prompt, but not make a whole girl disappear.

>>7012495
Wasn't there a discussion about how samplers are broken for X/Y above? Not sure if relevant, see >>7012150
>>
>>7012509
amazing efficiency but i wonder how good it is for high quality stuff
>>
>>7012458
You can't go wrong with Rin. Trust me. I've generated over 6,000 Rin's since I got my hands of SD.

>>7012461
>>7012479
Oh... wait, so the line break is actually intended and not some weird Copy/Paste malformation. Interesting.

So, like the other anon said, quality tags would by included alongside style, and.. are we sure about giving the pose before the description (which I assume includes the character, if you want it to be a specific one)?

One of my better goth Rins in return. I'm glad I came to this thread for help.
>>
>>7012515
>I'm glad I came to this thread for help.
/hdg/ is easily the best place for assistance and/or tips.

/g/ COULD be the best place except they're unhelpful assholes. /aco/'s proompting quality is basically still what it was back early September. /d/ learned some tricks early on that trickled back into here but I am not sure they're as active now. I remember some /d/-anon coming here and telling us there was no active /d/ thread at some point last week
>>
>>7012512
https://files.catbox.moe/r1gr8m.jpg

a X/Y plot run on anythingv3 with SDE
>>
>>7012526
>I remember some /d/-anon coming here and telling us there was no active /d/ thread at some point last week
Probably the same one who kept thinking there was no thread on /e/ and just didn't bother looking. The /d/ thread's a lot slower but it's still running.
>>
File: Tiddies.gif (3.91 MB, 512x512)
3.91 MB
3.91 MB GIF
>>7012361
There is good stuff waiting to be made without training. I did promise you guys I would make something KINO when I find the guide is in a more or less complete state but I lack feedback and people trying it out and asking questions. I hope some anons give the guide a try.
>>7012384
>>7012479
Thanks for spreading my screencaps anons. You guys are great!
>>7012526
Sorry to hear people haven't been helpful anon. It's an eclectic mix of characters on at certain times. I have learned much though and news hits there pretty quick. I miss the inpaint animator anon... ;_; That's all for now. Take care everyone! I hope you have a great day!
>>
>>
>>
>>7012535
holy shit
prompt pls
>>
>>
>>
>>7012479
this sounds interesting but I don't quite get it >>7010773 what are outer most tokens, left and right.. like how to visualize that in context?
>>
holy fuck lmao
>>
>>7012554
start and back of prompt
>>
>>7012552
that is super hot. would you be kind enough to share the prompt
>>
>>7012559
(best quality), nsfw, nude woman, black witch hat, green skin, pussy, orange hair, (((leather restraints))), light smile, tentacle vaginal
Negative prompt: lowres, bad anatomy, (bad hands), text, error, (missing fingers, extra digit, fewer digits), (extra arms), cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, depth of field, censored
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 12, Seed: 898621222, Size: 1024x1280, Model hash: 925997e9, Denoising strength: 0.67, Clip skip: 2, First pass size: 0x0
>>
>>7012554
Here's how I understand it, any anon with more technical understanding may jump and correct me:

masterpiece, best_quality, 1girl, blonde, anime style, blue bikini

When the AI reads this, it starts considering/arranging/painting, in order
1. It must be close to what I was taught consists a masterpiece
2. I must add a bikini in there
3. It must be similar to what I was taught can be described as best quality
4. I must add blue in there
5. It must have one girl
6. It must have blond hair

This is a slight problem if true because then the ideal prompt would have to be written in an unnatural way. We'd have to write the prompt the way it comes naturally to a human, then prioritize each aspect, then start moving things left and right.

Totally incidentally, has anyone seen the movie Arrival (2016)? I *strongly* recommend it... I don't know why, but it just jumped into my mind /s
>>
>>7012554
>>7012556
ok I'll try to monologue this for myself so I can understand

if I have... 7 random tags as example
>highres, flat color, 1girl, school uniform, doing laundry, sunset, squatting

does this mean that the three tags in the middle should be the ones defining the character, while the two tags at the start and end of the prompt are setting the scene so to speak?

..but then does that mean that some tags that you would generally use first on the left, you would then instead keep for later to use on the right side of the prompt
>>
>>7012569
>This is a slight problem if true because then the ideal prompt would have to be written in an unnatural way. We'd have to write the prompt the way it comes naturally to a human, then prioritize each aspect, then start moving things left and right.
exactly which is why the current version of CLIP is retarded
>>
>>7012462
cheers
>>
>>7012479
>>7012569
>>7012583
>>7012585
>>7012515
very interesting stuff, alas can man get to the bottom of this
>>
>>7012591
the anon that has been prompting catgirls in front of landscapes solved this like 3 weeks ago
>>
>>7012450
Getting any kind of fingering prompt to work is extremely hard
>>
I'm trying to do an X/Y plot, but when trying to use different samplers, it doesn't work (it just takes the one that is select by default), does someone knows why this happens? (This ran euler a on the 3 samplers instead of euler a, karras and ddim)
>>
>>7012600
it's broken you need to revert to an earlier commit
>>
File: Ee9_trZWkAEWYIp.jpg (30 KB, 1200x233)
30 KB
30 KB JPG
>>7012346
>>
>>7012601
Thanks anon, any previous than the latest commit or is some commit in specific that broke this?
>>
>>7012610
scroll up
>>
>>7012593
and yet
>>
>>7012593
~2500 years ago, Plato wrote that "our need will be the real creator" (nowadays: necessity is the mother of invention).

Truly, humanity's need for catgirls would inevitably reveal a pioneer among us.
>>
>>
>>7012373
>december
>wd 1.4
good one. Release date is being moved back another month as we speak so they can train on SD2.0
>>
Been testing out the DPM++ 25 2M Karras and hot damn it does some fine work. Shorter processing time than Eular and pretty amazing results (Even if bit random at times)


1girl, solo, ((masterpiece, highest quality, ultra detailed, highrez)), (petite young woman drinking alcohol in her apartment), twintails, strapless bra, open hoddie that show her midriff and bra, (android, cybernetics:1.4), robotic joints, (beautiful detailed hair, fair skin and extremely detailed clothes, beautiful face:1.3), (cyberpunk, neon lights, neon colors,:1.3), small breasts, pink tattoos, dark room, dim lighting, wide hips, thin waist, bed by a window, lips, (laying down), (hand between crotch, masturbating, fingering, nsfw:1.4), blushing, piercings, lots of dildos and vibrators on a bed, cameras
Negative prompt: multiple girls, lowres, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, extra fingers, mutated hands, (multiple penises), (disembodied), (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), animal ears, extra ears, fat, cross eyed, tanlines, (bad perspective:1.3), fat, dark skin, multiple legs, multiple breasts, men
Steps: 40, Sampler: DPM++ 2M Karras, CFG scale: 6, Seed: 4054731105, Size: 768x960, Model hash: 925997e9, Model: nai, Denoising strength: 0.65, ENSD: 31337, First pass size: 448x640
>>
>>7012646
whats the point of having multiple prompts in an ()? does it do anything special?
>>
>>7012612
Fixed it by applyng an open pull request, but thanks nonetheless.
>>
>>7012649
sorry i would have given the answer if i knew it but i didn't update because i heard about this bug so idk
>>
>>7012646
not to be harsh on you anon but we figured this out last month ago, you can still cut about half your steps as well
>>
>>7012650
don't worry, thanks to you I checked the github and then the pull requests, without your suggestion to check an older commit I wouldn't have known lol
>>
>>
>>7012647
gives them all the same amount of emphasis without having to put the ()s around each individually, that's about it
>>
Tried tha same promp with multiple Step size. Like 20, 30 and lastly 40. In my preference the last had the best detail, so just happened to post that one
>>
>>7012659
with dpm i see improvements up to 50 steps
>>
>>
>>
File: 1641893226518.png (618 KB, 512x768)
618 KB
618 KB PNG
>>7012511
Indeed. It's a mystery that's going to keep me up at night.

I do wonder, are you still able to generate the same two images on the latest commit? From the commit hash you posted earlier, it seems you also git pull'd to the latest hash like me, so it must have been a bit of time since the last time you generated those two.

I've tested a bunch of the other settings and they are all minimal effect to no effect such as changing eta noise. Narrowed down significant changes to clip skip (but that doesnt need to be changed for this model), hypernetwork (should be the exact same one so I doubt its that), and latent space fix for highres. I'm wondering about the above because webui often has commits that change how latent space fix works. I remember reading a github issue where someone said the a new commit change to latent space fix changed how an image looked like significantly when using the same options.

Also, although highresfix is a type of img2img, it looks like changing the upscalers for img2img has no effect on the latent space fix option itself. None of the settings above it have any effect on text2img besides the latent space fix option.
>>
>>7012663
I wasn't looking at the convo before so im NTA but if you are running highres.fix you will get different outcomes based on GPU
>>
ai kind of knows what a pentagram is but not really
>>
A new maid arrives
>>
>>7012474
>>7012427
>quantization
what's that and do I want it on 3070?
>>
>>7012660
>>7012646
>>7012650

20 CFG actually is quite good, but higher CFG might give more detailed results, as anon above mentioned (not sure if worth the processing time tho...)
Like 20 CFG is still really good~
Batch count 1,3,4 for this seed are pretty good, even if some in painting would be required:

1girl, solo, ((masterpiece, highest quality, ultra detailed, highrez)), (gothic lolita with big breasts striping off her clothes:1.2), sweat, (tattoos), (long eyelashes, make-up), dress big with boob window, vampire, seductive eyes and smile, (sexy, slut) vampire, (skindentation:1.2), long hair, hair ornament, (beautiful detailed hair, detailed intricate eyes, fair skin and extremely detailed clothes:1.3), piercings, cathedral, long finger nails, lips, corset, night, dim lighting, unholy, removing clothes, (cowboy shot:1.2), black and white, illustration
Negative prompt: multiple girls, lowres, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, extra fingers, mutated hands, (multiple penises), (disembodied), (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), animal ears, extra ears, fat, cross eyed, tanlines, (bad perspective:1.3), fat, dark skin, multiple legs, multiple breasts, men
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1145943222, Size: 768x960, Model hash: 925997e9, Model: nai, Denoising strength: 0.6, ENSD: 31337, First pass size: 448x640
>>
>not into feet
>generates image with bare feet
>likes image
>tries to make more images with bare feet
>refuses to work
>still think feet are gross but have to spend hours trying to make feet images to beat the machine
>finally get a good one
>foot is on the wrong fucking side
>>
I've cracked the code to good yuri! Just draw it lol... but this time I upload the right version
>>
>>7012685

yeah, you think women's feet are gross

if you're a retard
>>
>>7012685
>finally get a good one
>extra wide
>6 toes
????
>>
>>7012691
anime feet good
real feet bad
>>
>>7012690
clean those shit smeared fingers first
>>
File: file.png (188 KB, 384x424)
188 KB
188 KB PNG
Huh?
That's just a.i. bro not much I can there.
Still looks better than 99% of the fingers all the other retards generate.
>>
>>7012700
https://aibooru.online/
>>
>>7012694

nah, all good
>>
>>7012685
faggot
>>7012699
It looks terrible and obviously edited. Use cleaner lines when editing next time.
>>
>>7012700
Pixiv
>>
geeze, it's not like you even need to just only use tags, went and just described it randomly and my 1050ti just shat out this
>>
>>7012699
other anon is right that looks obviously edited cmon now
>>
>>7012707
if you have restore faces on you should turn it off
>>
Any more kuroko shirai hentai?
>>
>>
>>7012713
I love this anon he comes in like once a day and ask's if anyone made this character. the answer is no anon
>>
>>7012711
I didn't think I did but I suddenly found the option in the options and selected none. Thanks, I'm mostly just clicking random buttons and using random models.
>>
>>7012716
its how I learned anon, no worries. restore faces just has a very obvious tell and your anime results will be better without it.
>>
>>7012713
Okay I take back what I said, if you paypal me 30 dollars I'll make a dreambooth for kuroko shirai and gen you 30 images of her being raped by ugly bastards
>>
File: 00446-~1.png (291 KB, 512x512)
291 KB
291 KB PNG
>>
File: 3.png (3.18 MB, 1760x2400)
3.18 MB
3.18 MB PNG
Does anybody know if character hypernetworks can somehow be mixed with art style hypernetworks?
>>
>>7012713
do it yourself lazy
>>
>>7012725
but how anon?
>>
>>7012737
read the op like we all did 6 weeks ago
>>
>>7012737
>imagine being actually retarded
>>
>>7012739
Anon the feet breeding puts arent real.
>>
>>7012747
you are drunk
>>
Can anyone recommend good artists for sexy clothed characters? Bonus if they do slutty outfits and thickness.
>>
>>7012750
I like using the zanamaoria hypernetwork to make stuff like that
I used it for >>7012724
>>
>>7012750
a side idea you can put designer brand clothing in and you can get sexy dresses that way, there was a whole x/y someone made
>>
>>7012750
also adding navel to prompts can make the outfit have a cutout there.
>>
>>7012751
looks good, thanks

>>7012752
actually using that in my dynamic prompts somewhere but I think training over nai burns those out a bit
>>
The multiple subjects script doesn't work so well. The features bleed between subjects, the girls look pasted anyway, and the right girl was initially an orc.
>>
>>7012554
This anon got permanently banned >>7010773 , don't expect a reply from him
>>
>>7012763
>
>>
>>7012763
How would you know
>>
>>7012761
you can't just post that and not post prompt
>>
>>7012767
he's that anon and he's tired of your bullshit
>>
>>7012762
Honestly, even simply prompting everyone together works better. Not exactly as I wanted, and the clothing is mixed, but better anyway.
>>
>>7012762
isn't it supposed to build each girl separately how does it bleed? well glad I didn't take the time to try it
>>
>>7012778
It does a final img2img to blend everything together, and that's when it all goes to shit, I think
>>
>>7012778
it builds images then crops out the subjects from them then pastes them on the background image then img2img's them all together. it's no btter than photoshop except the ai is doing all of the copy pasting for you which honestly saves a lot of time
>>
>>7012778
Original image for the girl on the right in >>7012762
Before she is pasted
>>
>>7012783
right it does an img2img at the end, maybe the denoise is too high?
>>
>>7012784
Yeah, and with a low denoise it looks like this.
>>
Gape model is pretty great
>>
File: tmp2lv7tl6k.jpg (618 KB, 3072x1509)
618 KB
618 KB JPG
>>7012768
Its a custom model from random stuff I like, why I was asking for sexy clothes artists

>masterpiece, best quality, ultra-detailed, 1girl standing, elf, athletic, ornate ornamental patterned bridgeless bra, pleated skirt, lace trim, lace-trimmed skirt, high-waist skirt, earrings, happy, grocery store in background, sweat , steaming body , cum on clothes, by hero neisan, masami chie, pinkdrawz
>>
>>7012785
Ok, I think I might be a genius. It seems the trick is to use the whole-image prompt from >>7012775 for the final img2img
>>
>>7012785
I know its not what you wanted anon but I'd fuck all three of them
>>
>>7012788
they still look pasted on but it's nothing like before
>>
>>7012788
There are glasses on the orc and random errors in the background, and some hair/eye colors are swapped. But I think this might be useful, especially if you don't push it with different skin colors, very long hair, poses and so on
>>
>>7012714
i'd like to give this a go if you got the prompt
>>
currently on a mission to 10/10ify a less than optimal illustration with inpainting................
>>
>>7012803
Changing the poses makes this a bit more lively, still looks like VN characters on a background
>>
File: 00135-3544018054-toes.png (291 KB, 512x512)
291 KB
291 KB PNG
>>
>>7012800
(best quality), nsfw, nude woman, blue witch hat, pale skin, dripping pussy, long dark teal hair, hair over eye, shy, womb tattoo, in bath, orgasm, trembling
Negative prompt: lowres, bad anatomy, (bad hands), text, error, (missing fingers, extra digit, fewer digits), (extra arms), cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, depth of field, censored
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 12, Seed: 4093803585, Size: 1024x1280, Model hash: 925997e9, Denoising strength: 0.67, Clip skip: 2, First pass size: 0x0

Good luck, it was being a real bitch giving me good anatomy with that one
>>
>>7012806
bro, anatomy doesn't work like this
>>
>>7011727
No, i'm at work and it's currently training, i managed to finish all my tags, added the commas, had to recrop alot and change out some useless images but it's finally ready, will have something to show in a few hours. I'm saving up for an Nvidia which is gonna be about 500 sheckles so the process should be faster in maybe 3 weeks.

Also, here's my pixiv

https://www.pixiv.net/en/users/9936494
>>
what prompts are there for kuroko shirai?
>>
>>7012721
Not what I asked for retard.
>>
>>7012826
>>7012828
read the op
>>
I made a post here >>>/g/89958325 of a modified version of the image browser:
>https://github.com/yfszzx/stable-diffusion-webui-images-browser

I added 3 tabs, modified the delete function, and renamed the Favorites tab to "Cookbook"
>3 new tabs
First off you need to make 3 directories in the output folder: promptref , reference , trash
[Hotpot] is for the output of shared prompts. Also added a "move to Hotpot" button
[Heirloom] is for shared images containing metadata. No 'move to' button
[Trashed] is for 'deleted' images using the delete button since the interface can act wonky and sometimes delete the wrong image.

Favorites was change to [Cookbook] just to match the theme.
Its made for unix based systems. I don't know if it'll work properly for windows in its current form, but you can try it. Or even modify it again if you want.


images_history.py
>https://pastebin.com/PWEP10zL

this is the diff output if you are wondering what has been changed
https://pastebin.com/bj1yjXty
>>
>>7012828
>offer to make you the hentai you keep asking for
>thats not what im asking for
okay anon whatever you say
>>
>>7012826
Ask for Kuroko shirai hentai and go from there
>>
>>7012792
i coulnd't even get the thing to do anything, it would render the background and then fail out
>>
a 2080 super is too low to dreambooth right? I really, desperately want to train a model on this artstyle
>>
>>7012831
thanks for this anon.
Never realized how shit the extension's code actually is, I'm tempted to rewrite it from the ground up.
>>
how do I share my embedding anonymously, mega.nz?

I am trying to cherrypick which embedding is best to upload, is it possible for an embedding to be overtrained in the span of 100 steps at a learning rate of 0.0005 ? I.E. 25500 to 25600
>>
>>7012831
Fancy seeing you here.

I'm glad you learned how to do relative paths. Did you finally take a look at modules/paths.py?

>you need to make 3 directories

Why not just

try:
os.mkdir("./outputs/promptref")
os.mkdir("./outputs/reference")
os.mkdir("./outputs/trash")
except Exception as e:
print(e)

Though I think it'd be best if we actually use the log directory for promptref and reference...
>>
>>7012861
Actually if we want to be smart we can also just do

if not os.path.isdir(path):
. os.makedirs(path)
>>
>>7012861
>Why not just [...]
Thats simple,
it because I don't know python. I'm just reading the code and learning what needs to be changed to make it do what I want. Instead of starting off with an understanding of what my options are and what the language can do.
>>
>>
>>7012872
Hey this is nice!
Could you put 'shiny skin' or something in the negative? That's pretty distracting. Otherwise, hot damn please, more like this.
>>
>>7012867
No offense meant Anon, but I was moreso asking why not just have the code make the directories. If you don't know how to do that (or if it's possible), a quick search would help you.

The less work your user has to do to get shit working, the better.

Also it's been a while since I touched Windows, but I think it's fine if your paths have / instead of \\, since the OS module is smart enough to handle that or soMETHING. Never dug into it.
>>
File: 00857-708668537.png (2.19 MB, 1024x1536)
2.19 MB
2.19 MB PNG
your vat prompt was pretty cool anon, its inspired me to try something
>>
>>7012855
could try anonfiles
>>
>>7012855
catbox?
>>
>>7012880
thank u

she seems kind of nonplussed
>>
>>7012894
I was thinking the same thing. That face does not match the situation she's in. A little inpainting fixes that of course, but like, cmon.
>>
>>
>>7012896
while we're on the subject i discovered that expressions do indeed work fine under masks as long as you weight the mask enough so it doesn't draw a mouth instead
>>
>>
>>7012902
Do you have the prompt for this? This pose + creampie is exactly what I'm trying for.
>>
Is stenography against the rules on this site?
It would help get around site removing exif data if there was an extension that added buttons to the PNG Info tab that allowed you to encode and decode the parameters into the image.
I remember a nasty fiasco on /a/ a few months ago so I'm not sure if it's strictly off limits now
>>
File: 00181.png (1.45 MB, 960x1280)
1.45 MB
1.45 MB PNG
>>
>>7012907
Mmm Danone.
>>
>>7012906
What kind of fiasco are we talking about? I'm curious now.
>>
>>7012907
someone should tell them they got bird poop on their faces.
>>
>>7012906
> stenography
You mean steganography? It's kind of a gray zone. You will need a 3rd party app to write and read hidden information and theoretically decoded data can have a rootkit or virus. So this will count as a distribution of harmful programs. I am not sure it is possible to remove steganography data from image without re-encoding it, so hard to tell what 4chan can do to block it.
>>
>>7012640
haru didn't say that though
maybe few days but he said he was gonna train both 1.5 and 2.0
>>
>>7012904
here you go, i got inspired by this anon >>7012306

nsfw, bokeh, (highres:1.2), (ultra-detailed:1.2), (oil painting:0.9), (sketch:1.2), (extremely detailed CG unity 8k wallpaper:1.2), soft lighting, HDR, 1girl, facing viewer, thigh focus, close-up, (lying, on back:1.2), (spread legs:1.2), (((arms up))), (((legs up))), shiny skin, maid, (dark skin:1.3), black hair, short hair, swept bangs, ponytail, detailed pupil, beautiful eyes, detailed eyes, expressive eyes, red eyes, closed mouth, angry, (blush:0.5), perfect hands, medium breasts, nipples, (breasts out:1.5), wide hips, maid headdress, black gloves, bowtie, black dress, short sleeves, apron, black thighhighs, garter straps, wedding ring, skindentation, no panties, (detailed pussy, pussy:1.3), (cum in pussy, overflow:1.2), (sweat), skirt lift, (((on the floor and on a stage inside a luxury mansion))), dynamic angle, looking at viewer,

Negative prompt: lowres, bad anatomy, bad hands, mutated hand, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (((idolmaster))), genshin impact, touhou, fate \(series\), colored pencil \(medium\), out of focus, glowing eyes, (wedding ring:1.1), (((multiple views))), (((bad proportions))), (((multiple legs))), (((multiple arms))), pointy ears,
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 3299858302, Size: 960x1280, Model hash: 38c1ebe3, Model: Anything-V3.0-pruned-fp16, Denoising strength: 0.65, Clip skip: 2, ENSD: 31337, First pass size: 448x640
>>
File: 00182.png (1.6 MB, 1280x960)
1.6 MB
1.6 MB PNG
>>7012913
livin' that yogurt life
>>
>upscale latent space
actual game changer but what denoise do you keep it on? Alot of images come out blurry or with extremely detailed hair that doesn't look so good. Was using 0.5 for this
>>
>>7012932
Sharing is caring!
>>
>>7012933
0.5-0.7 depending and you will need to adjust your tags, at least use (detailed:1.2)

and you got satoko's eye color wrong
>>
>>7012894
yeah I kinda had a plan for it but I'm still working on it but I'm starting to think this might not be the ideal canidate
>>7012896
I'm indifferent, the reason being it could also be a situation they are familiar with. for example, she could have been in this suspended state for a long period, to which something like a insertion could be a long, ongoing, and uncontrollable situation. maybe its not fucking her at this moment. the truth is I want to animate it for a little with a simple sort of pulsing, (she would bob in the water) and maybe some bubbles, I just wanted to experiment with the idea.
>>
>>7012938
>wrong
Nah, was just aiming for witchtoko sea of fragments Sotsu aesthetic
>>
>>7012932
P-prompts...?
>>
>>7012943
anon her eyes are red not gold
>>
>>7012915
there was cp encoded in /dbs/ threads that went unnoticed for an unspecified amount of time
>>7012924
I juryrigged a working proof of concept using someones years old public code I grabbed from github but I'm worried that it'll be flagged if i even try uploading any encoded images lmao. I'm just gonna leave it be and go back to prompting
>>
>>7012945
I am but a slave to the machine's dice.
>>
>>7012948
inpaint
>>
>>7012947
This is why we can't have nice things...

>>7012948
sotsugou satoko is a shitoko. That considered, those look very nice. Prompts?
>>
Detailed's kind of a pain cause it influences AI's ability to draw Satoko consistently.

>>7012949
I am a firm believer of quantity for the most part, inpainting a picture is too much of a hassle usually and it just never ends once you begin.

>>7012950
masterpiece, best quality, {{houjou satoko}}, loli, {black hairband}, {{{blonde hair}}, red eyes, short sleeves, puffy sleeves, solo, short hair, higurashi no naku koro ni, {{green dress}}, {{{straight_hair}}}, white sailor collar, {{{yellow neckerchief}}}, intense angle, depth of field, cinematic lighting, serious, dynamic angle, floating glass, shattered glass, lens flare, glass fragments, magic, space, {{masterpiece}}, smirk, highres, {{masterpiece portrait}}, looking away, dynamic pose
Negative prompt: longbody, 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, deformed, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, messy drawing, broken legs, fat, hands inside body, extra leg, third leg
Steps: 40, Sampler: Euler a, CFG scale: 13, Seed: 3502266915, Size: 1088x1088, Model hash: e043dfc5, Denoising strength: 0.5, Clip skip: 2, ENSD: 31337, First pass size: 512x512
stole off AI booru, added my own negatives and switched up the model to strawberry, as far as the uncanny ones go - rest are berrymix, nutmeg or standard NAI
>>
>>7012947
>there was cp encoded in /dbs/ threads that went unnoticed for an unspecified amount of time
the fuck
>>
>>7012953
Thank you for the prompts! Also, you could just use gimp or something to edit the eye color with hue & saturation and then inpaint them with very low denoising.
>>
>>7012956
There's no need for all that stuff if it's this consistent - I'd rather just add golden eyes to neg or something. The only things I wanna inpaint are unusual hair colors sometimes but I'd have to boot up photoshop for that shit which I don't even wanna bother learning. Could also carefully paint the hair in paint and img2img till it looks consistent but as I said that's just hours of work.
>>
what are the tips for making the ai draw a certain anime character?
I mean I've been using the tag you'd usually be using on boorus for her but there's always something missing or it's really imperfect.
>>
>>
>>7012947
> there was cp encoded in /dbs/ threads that went unnoticed for an unspecified amount of time
Well, that's another way of using it. Though the main principle of steganography is that you can't prove that there's a hidden information inside another information. So technically any image that is posted unchanged can have CP, yours included. We just need the right key to find it there.
>>
>>7012962
>char series
>char name
>describe hair/eyes
>describe outfit
I found out adding "detailed" or "ultra-detailed" either helps or makes it less likely, berrymix or nutmeg sometimes get a char better than base NAI aswell so it's worth checking a diff mix
>>
>>7012959
Some SFW is fine but at least occasionally post some hentai (but not of Satoko, that's only allowed on /b/).
Also {{ }} is only used on NovelAI's website, WebUI uses (( ))
>>7012962
Give an example of a character
>>
>>7012969
>>7012962
oh also adding other characters from same series to negatives - for example trying to prompt fate charas without adding Saber and Tohsaka into negs isn't gonna end well.

>>7012964
>>7012947
wait is that why these threads were so popular? I heard it was bots in there but nothing about CP. People still use thirdeye/pngembed sometimes but I dunno if that'd be helpful for prompts/img data and iirc it's not allowed, but you're unlikely to get banned either
>>
>>
File: 1645377905844.png (803 KB, 768x1024)
803 KB
803 KB PNG
pretty happy with this batch. not smart enough to figure out how to get her nude, but frankly im in awe with how clean some results can look.
now to find the time to make a chamchami hypernetwork
>>
>>7012962
> what are the tips for making the ai draw a certain anime character?
Train a hypernetwork.
Seriously, lots of artists draw same character in totally different styles which you may not like, and when model learns it creates knowledge about something in a middle of all those ssamples, which looks similar to original character but not exactly like it. You need a dedicated model trained on handpicked set of images to get what you really need.
>>
>>7012977
chonky, post more.
>>
>>7012973
im aware, I just didn't bother changing the prompt I stole since it worked. Also weirdly enough {} seems to work similarly to (), since it did actually affect her hairband appearing more often when I added it. No clue what the hell's up with that.
>>
File: 1641193331246.png (830 KB, 768x1024)
830 KB
830 KB PNG
>>7012979
have one or two other ones that are different enough to worth being posted. was recycling the same seed and playing with hypernetworks/specific phrases in the prompt
>>
>>7012975
> wait is that why these threads were so popular?
Ni, they're popular cuz AI generated images are fun and hype right now. But I cannot really prove that there's no hidden CP in any of images posted in this thread, and noone can - that's the beauty of true steganography.
>>
>>7012969
>>7012973
I've been trying, name, series, hair, clothes.
with (()) and without
A hundred tries just mashing buttons, models and hyperthreads.
best I got is this,
char is supposed to be Yuzu Hiiragi from yugioh's arc v, hiiragi_yuzu on boorus

>>7012978
I'm on a shitty little 1050ti so that's out of the question.
>>
More maids will be creampied at a later time.
>>
>>7012987
>Yuzu Hiiragi
>557 images

You're fucked Anon. You're better off taking an image off of her from Danbooru and using deepbooru interrogate to try and get her tags, then filter through redundant ones from there.
>>
>>7012987
Well, that sucks.
The, first you need to use character's name tag and assign it highest weight, then if it doesn't have features you want - describe them but with lower weight. It will try to generate details based on a most-weighted tag - the character you prompted. This method has a big flaw though - if character is not really popular and had small training set you won't get reliable good result, maybe by luck.
>>
>>7012986
oh, I thought you meant the dbs threads on /a/, seems it was something else entirely, my bad.

>>7012987
alot of characters just aint possible without a network, they need to be moderately popular and have designs that the AI wont struggle with - if they're too generic it's bad, if they're too hard to make it's bad. Sometimes you'll get lucky with a low popular chara too but dont count on it. Can always inpaint as last resort or img2img anime screenshots etc.
>>
>>7012930
Thank you! I mostly use F111 for realism but I'm on here because this board is way more advanced in its understanding of how this tool works. It doesn't seem like F111 has any creampies in its training set so I'll probably have to do a merge.
>>
File: 1 (18).png (497 KB, 640x640)
497 KB
497 KB PNG
>>7012991
Bullshit, I've made Sansei Muramasa way back when I started out, with custom clothes and even lolified and she doesn't even have 500. It just works sometimes.
>>
>>7012993
I wasn't there far these dbs theads so I don't know, maybe they were popular cuz CP was shared there and people knew how to retrieve it.
>>
>>7012995
Actually 500 images is a lot. Hypernetworks are usually trained on a set of 50-100 and this is enough for them.
>>
>>7012995
You know what, I'll concede to you. Just that in my experience it sucks far horse cock in those cases.
>>
File: 09957-2022-11-26.png (971 KB, 896x1152)
971 KB
971 KB PNG
>>
wanna know the best reason to share prompts?

If you're horny all the time and have specific stuff that gets your hog hard, you can make a good prompt of it, share it out to the anons, and let them itterate on it and post stuff.

then you get more stuff you wanna jerk off to.

(wet raining),(nsfw), (fullbody), (realistic, [anime]), (3d:0.3), dramatic lighting, sitting on intricate throne, ((masterpiece)),(quality),(highres), tall harpy, [[covered abs]], ((chrome power armor|latex rubber bodysuit) mechanical arms, prosthetic legs), (cameltoe), long white hair hair ponytail, [guilt], small breasts, covered navel, eyeliner, eyelashes, perfect face, beautiful nose, detailed pupil, beautiful eyes, detailed eyes, wool eyes, perfect lips, outdoor shrine pillars vines background, sunbeam, light rays, [boku no hero academia], (dramatic lighting:1.4)
Negative prompt: lowres, bad anatomy, bad legs,bad hands,error, missing fingers,text,extra digit ,fewer digits,cropped, worst quality ,low quality,normal quality, jpeg,artifacts,signature,watermark, username, blurry, bad feet,artist name,bad anatomy,bad hands,bad body,bad proportions, worst quality,low quality
>>
File: musashi.png (1.49 MB, 1152x768)
1.49 MB
1.49 MB PNG
>>7011541
Hotter, mind sharing the prompt for your pic?
>>
File: tn3324sa230s0.png (1.64 MB, 1024x1536)
1.64 MB
1.64 MB PNG
>>
>>7013006
Whoa, I really like the tentacles clamping on the nipples. Was that deliberate?
>>
>>7013011
nope, happy accident!
>>
>>7011344
bro wtf is up with her hand
>>
>>7013030
Damn, I was definitely hoping for more of this.
Do you have wrapped or something like that? Looks like it's going for a gagging kind of thing
>>
>>7012855
I guess my embedding failed, tried out the different steps but does not look like stuff from
sincos

in fact it looks like regular old NAI
>>
File: tmpdz0n41th.png (2.52 MB, 1280x1920)
2.52 MB
2.52 MB PNG
>>7013034
not quite wrapped, but...
>>
>>7013045
I meant 'wrapped' in the prompt because it looks like it was choking her. Either way, these are great! What model is this?
>>
File: 1643558417361.png (1.11 MB, 896x1152)
1.11 MB
1.11 MB PNG
>>7013003
i appreciate your transparency, i think i'll iterate on it
>>
>>7011738
Holy shit. I've been having this problem with raspberry and other mixes for a while, thanks a ton.
>>
>>7013050
With VAE manually selected
>>
>>7013048
titterate
>>
Are hypernetworks better at art styles than dreambooth models?
I feel like the dreambooth models for art isn't really that good compared to hypernetworks + AnythingV3.
>>
File: file.png (939 KB, 704x768)
939 KB
939 KB PNG
>>
File: Final.png (1.51 MB, 1088x2048)
1.51 MB
1.51 MB PNG
Yo like pee and stuff yeah fuck d
>>
Remember that multiple images extension that was linked yesterday? I think anon was right, it was compromised, My python sends large chunks of data somewhere.
>>
>>7013081
could this be related to the anons that got their systems compromised?
>>
>>7013081
which multiple images extension? That's not great
>>
>>7013090
I've already deleted it. One that allowed to create multiple foreground images and merge them onto one background image.

>>7013089
I don't know, maybe.
>>
>>7013081
specifics? show what made you think this was the case
>>
File: 20221120203956- .png (1.22 MB, 1024x768)
1.22 MB
1.22 MB PNG
>>7013081

Confirmed traffic is coming from this extension. Multi-Subject Extension is NOT good to use.

>Pic related.
>>
>>7013089
No, my system was comprised by an unrelated incident. I downloaded a 4year crack for l2d editor because I've been experimenting with the idea of animating one of my good prompts.
>>
>>7013081
>>7013099
I took a gander at the source just now and it's very simple. The only traffic that should be incoming (not outgoing) from the script is to clone the MiDaS repo, and to download the depth model. You should be providing more evidence what exactly else may be happening before throwing these claims out there.
>>
>>7013105
Here, I found a bit more info. It's not necessarily the extension's fault, it's the MiDaS repo being used that's causing it, and it's not something nefarious. The dev is going to fix it. https://github.com/Extraltodeus/multi-subject-render/issues/14#issuecomment-1327962378
>>
>>7013081
This anon thinks the local to local data is a hack
>>
File: 20221120203867- .png (651 KB, 576x768)
651 KB
651 KB PNG
>>7013105
>>7013111

>>7013105
>>7013111

I don't want to spread Fear Uncertainty or Doubt, but my router security tripped and kicked me off the network with this extension. I won't use it anymore until I get better OpSec.
>>
>>7013122
Well you were clearly right about it being an issue >>7013111, it just wasn't malicious
>>
>>7011312
There is more. There is just shitload and no one have energy to bother really.
For example there is this https://huggingface.co/Fansy/poison Finetune on Anything-3.0 to make anime more realistic, but still be nice and consistent.
I do recommend it.
>>
Doing some cleaning and organizing, mostly just deleting redundancy. Got a chuckle out of this.

>"See, I have a sense of humor? 'Honk honk'."
>>
>>7013122
>router security
Removing this and replacing it with Common Sense would be an excellent step in getting better OpSec. Are you running deep packet inspection on your own LAN or something crazy?
>>
>>7013116
Skindentation so intense that nipple looks like salami.
>>
Does SD v2 good?
>>
File: emad.jpg (28 KB, 324x405)
28 KB
28 KB JPG
>>7013143
sd v2 does not good sir
please do the needful and fine tune the model with anime bobs and vegene sir
>>
>>7013143
try asking that question again correctly
>>
>>
>>7013153
Are SD v2 did?
>>
why has no auto1111 webgui updated with sdv2. we fall behind sirs
>>
>>7012947
Is this about that extension called Third Eye or something?
Completely forgot about it till now
>>
>>7013153
Drug whore. Feed her a bunch of pills, anon.
>>
Anyone figured out a good prompt setup for juicy_details tag on boorus? Basically cutting out the head / other parts to focus on core body.
>>
>>7012989
Prompt for these ones?
>>
>>7012989
Do you have a prompt for this one?
>>
wait tentacles are allowed in /h/ now???
>>
>>7013047
oop, passed out for a bit. nope, it's just ((sex) with tentacles). I believe that model was Rando with j.k. hypernet. it was before i turned on the option to save prompts along side images in text so i'm not 100%
>>
File: 00185.png (1.29 MB, 1040x1040)
1.29 MB
1.29 MB PNG
i'm only posting this because of the ferocious sideburns
>>
>>7013218
no, its not. i got a warning for it.

sorry mods!
>>
File: tmp72k99rzi.png (1.61 MB, 1024x1536)
1.61 MB
1.61 MB PNG
>>
>>7013226
this is every vtuber when the stream ends
>>
>>7013222
found my new haircut
>>
File: tmpo08ez3f9.png (1.74 MB, 1024x1536)
1.74 MB
1.74 MB PNG
>>7013227
and me when proompting!
>>
>>7013228
Kek, when you go to the barber show him that exact picture
>>
>>7013222
Me on the left doing cum-based magic tricks for the boys while we wait for our turn
>>
>>7013235
the forgotten sixth element
>fire
>earth
>metal
>water
>wood
and
>cum
>>
>>
>Hey anon, what did those guys do to you anyway?
>Doesn't matter, they won't bother you anymore sweety
>>
I've trained a custom model to do a few different art styles.
https://rentry.org/EverythingModel
>>
>>7013242

I feel like I've seen her before. Or it's just a coincidence and I'm flattering myself.
>>
>>701325