[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
/g/ - Technology


Thread archived.
You cannot reply anymore.




Old: >>88772350

>Local install
GPU: https://rentry.org/voldy | https://github.com/AbdBarho/stable-diffusion-webui-docker
CPU: http://rentry.org/cputard

>AMD GPU Guides
Native: https://rentry.org/sd-nativeisekaitoo
Docker: https://rentry.org/sdamd
Onnx: https://rentry.org/ayymd-stable-diffustion-v1_4-guide

>M1 Supported Fork
https://github.com/lstein/stable-diffusion

>Cloud Hosted Install
Colab: https://colab.research.google.com/drive/1kw3egmSn-KgWsikYvOMjJkVDsPLjEMzl
Paperspace: https://rentry.org/865dy

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

>Krita Plug-ins
https://github.com/sddebz/stable-diffusion-krita-plugin
https://www.flyingdog.de/sd/en

>Other Guides
Inpainting/Outpainting: https://rentry.org/drfar
Upscaling images: https://rentry.org/sdupscale
Textual inversion: https://rentry.org/textard
Training: https://rentry.org/informal-training-guide

>In-Depth Usage Info
/g/ wiki: https://wiki.installgentoo.com/wiki/Stable_Diffusion
SD Akashic Records: https://github.com/Maks-s/sd-akashic

>Tools & Resources
Big index: https://pharmapsychotic.com/tools.html
Artist styles: https://rentry.org/artists_sd-v1-4 | https://www.urania.ai/top-sd-artists | https://sgreens.notion.site/sgreens/4ca6f4e229e24da6845b6d49e6b08ae7
Prompt sharing: https://lexica.art/ | https://krea.ai/
Dataset: https://rom1504.github.io/clip-retrieval/ | https://laion-aesthetic.datasette.io/laion-aesthetic-6pls/images

>Finetuned Models
https://rentry.org/yrpvv
https://github.com/eyriewow/merge-models

>Related
/h/ /sdhb/: >>>/h/stable
/vg/ /aids/: >>>/vg/aids

Official: discord.gg/stablediffusion
NSFW: discord.gg/2ZhBzDGf8u
Training: discord.gg/xVsyrmhQWS
>>
File: outpaint.png (341 KB, 1536x404)
341 KB
341 KB PNG
>>88773678
Anyone have any tips for getting good results from outpainting or just expanding an existing image in general? Seems like a total crapshoot if the resulting expansion is good enough to work with, and even if it is good, it seems like inpainting to fix issues always gives fucky results
>>
>>88773687
did you try experimenting with seed resizing?
it's in voldy's webui under extra
>>
File: REM.png (1.55 MB, 1024x1024)
1.55 MB
1.55 MB PNG
Everyone around, love them, love them
Put it in your hands, take it, take it
There's no time to cry, happy, happy
Put it in your heart where tomorrow shines
Gold and silver shine
Shiny happy people holding hands
Shiny happy people laughing
Negative prompt: ((dim)), ((unhappy)), ((frowning))
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 101950610, Size: 1024x1024, Denoising strength: 0.7
>>
File: file.png (559 KB, 512x512)
559 KB
559 KB PNG
>>
>>88773687
You are trying to outpaint the one region AI is the worst at representing, which are hands. Of course it's going to be a problem.
>>
How much is the new waifu diffusion model being delayed? I heard they changed the target to 600k or even 2 million images
>>
File: grid-0002.png (2.3 MB, 1664x1280)
2.3 MB
2.3 MB PNG
first for Touhou girls in Renaissance Painting style
>>
File: dessert moon.png (361 KB, 512x512)
361 KB
361 KB PNG
1000 hours in Stable Diffusion
>>
>>88773707
stable diffusion really can only do 1 face per image/seed, huh?
>>
>>88773717
Hopefully, forever.
>>
File: shack.png (279 KB, 512x512)
279 KB
279 KB PNG
Frogs ONLY past this point
>>
>>88773723
I had a lot of difficulty making actually accurate pictures of touhous without obliterating the picture with high CFG inpainting or rebaking. In particular what was pissing me off is that invoking Reimu's name gives her red eyes, every time.
>>
How taxing is this on the GPU? I've got a 3060 so the vram is finally to some use but I don't want to completely run it dry. Thinking of starting some big batches while i go to the gym is there some idea as to how big count/size should be?
>>
Retard are you okay
are you okay?
>>
>>88773702
Interesting I hadn't noticed that before. How does it work?

>>88773714
I understand the limitation on hands. I'm more focused on the torso area, the blending of the shirt, and touch-up inpainting not really working.
>>
File: file.png (626 KB, 512x512)
626 KB
626 KB PNG
>>
>>88773748
It runs your GPU at 100%. According to your manufacturer, that should be fine. They're usually correct.
>>
>>88773749
was for
>>88773722
>>
File: file.png (669 KB, 827x787)
669 KB
669 KB PNG
>>88773751
>How does it work?
image related
I think it doesn't work well with every sampler though. You need to experiment.
>>
>>88773758
I don't understand if this is supposed to be ironic
>>
File: 1663945787499324.jpg (107 KB, 1657x463)
107 KB
107 KB JPG
>>88773749
ftfy, you welcome.
>>
File: girl.png (406 KB, 512x512)
406 KB
406 KB PNG
first one to guess what profession (like "police", "witch") I used in the prompt to generate this girl gets a (You)
>>
File: download (54).png (568 KB, 512x768)
568 KB
568 KB PNG
The mistakes are part of the fun
>>
File: 1663915046291042.png (374 KB, 512x512)
374 KB
374 KB PNG
Keep walking coomers, this is a frog man's board.
>>
>>88773776
Stop samefagging
>>
>>88773762
cry harder.
>>
>>88773785
I like this
>>
>>88773788
you are literally the samefag.
>>
>>88773772
that is super fucking cool?
>>
File: rtjuer5456j.jpg (126 KB, 1710x576)
126 KB
126 KB JPG
>>88773776
ftfy
>>
File: genisis.png (518 KB, 512x768)
518 KB
518 KB PNG
I'm ignoring this mentally ill loon
>>
>>88773744

I use CFG 10, 20 Steps Euler A or Euler if that helps. Putting them in a painting and specifying (((touhou anime))) can help a little too. Also use negative prompts!! Like ((no red eyes))

hope this helps
>>
>>88773772
I'll test this out thanks anon.
>>
File: Screenshot (69).png (527 KB, 1920x1080)
527 KB
527 KB PNG
>>88773678
>me
>look at stable diffusion threads seeing people make amazing looking stuff with
>happy that /g/ finally grew up and instead of crying about distros they are doing actual programming
>finally found my place, look up the guides and start working with it
>turns out, you niggers literally just enter prompts into a textbox you dont even write code
>it's like making a thread about how to use google image search and acting like it belongs on /g/

i knew it, i fucking knew it. i don't know why i left myself be fooled again thinking /g/ is not just a bunch of uni-freshmen and lazy hacker wannabees anymore.
>>
>>88773819
she's not even got lungs she's a head attached to an ass kekw
>>
File: Blood_Pearl_2.jpg (660 KB, 1535x1531)
660 KB
660 KB JPG
>>88773817
stop wasting image slots kiddies
>>
>>88773785
How do anons make those? I tried img2img but it either wipes the original or just adds a bit of an overlay.
>>
>>88773832
So it's our fault you can't read
Got it
>>
File: you.png (279 KB, 512x512)
279 KB
279 KB PNG
>i knew it, i fucking knew it. i don't know why i left myself be fooled again thinking /g/ is not just a bunch of uni-freshmen and lazy hacker wannabees anymore.
>>
File: unnamed.jpg (54 KB, 900x900)
54 KB
54 KB JPG
>>88773817
> I changed the theme, he will think I'm not the samefag

ok brainlet
>>
>>88773832
>It's a "you don't do anything you are lazy no true artist"-episode again
Yeah we know.
So what?
Shut up and spam funny generated images or piss off.
>>
File: caveman knowlage.jpg (24 KB, 540x625)
24 KB
24 KB JPG
Could someone help me out with installing the proper Pytorch version that works with cuda on my 1080 with the conda on windows? I've already spend half a day fixing it up with dozen of different alternative combinations and constant installing and uninstalls of dependencies and environments.
Even the official pytorch local installation tutorial stop working for some reason and neither is piggybacking update requirements from the yaml in Voldy repo.
>>
>>88773832
admittedly a lot of us aren't doing any sort of coding yes(i'm barely an exception)
but that's mostly because it's a few people doing the heavy lifting on making this work as well as it does
>>
>>88773867
nope. if you run conda and then activate a named environment and run "pip install -r requirements" i think? it'll work out a version or its FUCKED and you should fetch again.
>>
File: explosive.png (612 KB, 512x768)
612 KB
612 KB PNG
>take generic coomer prompt
>adds "suicide bomber"
>>
File: download (47).png (601 KB, 512x768)
601 KB
601 KB PNG
>>88773846
Nice boat
>>
>>88773832
Don't know what you expect from this thread besides guides on how to set it up or people posting their art, and settings for CFG and prompts. you're just retarded
>>
>>88773678
I can't keep doing this all day.
>>
>>88773832

I can't code shit, but learned a ton, can't this be about learning and having fun?
>>
>>88773850
good prompts, luck and 0.5-0.8 denoise
>>
>>88773883
and boom goes the dynamite
>>
File: 546515300.png (422 KB, 512x512)
422 KB
422 KB PNG
>>88773832
you gotta be pretty dumb to think random people on 4chan are capable of building this type of AI
>>
>>88773779
priestess
>>
File: index.png (1.06 MB, 1024x1024)
1.06 MB
1.06 MB PNG
Does anyone know a fix for when you generate a 1024x1024 image and the program creates doubles like this?
>>
File: 19349857854763496.png (281 KB, 576x512)
281 KB
281 KB PNG
>>88773850
You have to be imaginative with it. Just throw a bunch of prompts at the image and see what the AI does with it.
>>
File: girl2.png (377 KB, 512x512)
377 KB
377 KB PNG
>>88773905
Nope. Here's another hint.
>>
File: 999.jpg (238 KB, 509x509)
238 KB
238 KB JPG
so what ya proompting today anons
me? just ancient records of biomech tech
>>
File: vr.png (399 KB, 512x640)
399 KB
399 KB PNG
>>
File: file.png (249 KB, 512x512)
249 KB
249 KB PNG
https://www.youtube.com/watch?v=GnrwM7vFn_U
all my prompts theme song
>>
>>88773773
>>88773758
I'm not being ironic or joking. I'm completely serious. Running your GPU at 100% TDP shouldn't destroy it in the short term.
>>
>>88773886
then the 5 constant stable diffusion threads that are taking up catalog space shouldnt be on /g/.
this is like those coomer software general threads all over again
>>
art styles are beginning to merge
>>
>>88773832
I don't believe you.
You even use the waifu model.
You are a shitposter.
Anyways, shut up and post more generated images
>>
>>88773832
>happy that /g/ finally grew up and instead of crying about distros they are doing actual programming

That's on you, nigger. This is /g/ - Technology, not /g/ - Programming.
>>
File: 509438346.png (394 KB, 512x512)
394 KB
394 KB PNG
>>
>>88773867
just use their conda environment
>>
File: medievil_girls2.png (2.15 MB, 1024x1536)
2.15 MB
2.15 MB PNG
>>88773937
>>88773832
>>88773832
>>88773832

>HE HAS 6GB OF VRAM
LAUGH AT HIM. LAUGH!
>>
File: tv.png (442 KB, 512x640)
442 KB
442 KB PNG
>>
https://twitter.com/anton_lozhkov/status/1572906509335498752
https://discuss.huggingface.co/t/decoding-latents-to-rgb-without-upscaling/23204
>Turns out that a simple linear combination of #stablediffusion VAE's 4 latent channels can produce a viable 64x64 image without running the VAE decoder
would this be useful for you anons?
>>
>>88773952
This.
>>
>>88773959
>unzips pants
>see face
>zips pants again
>>
>>88773832
My programming expertise peaked with making a linked list implementation for an sdl thing i never finished. I can't do more than that so prooomting it is.
>>
File: 5094383401.png (304 KB, 512x512)
304 KB
304 KB PNG
>>88773937
>After all, why not? Why shouldn't I keep it?
>>
File: 1663946735609-1.png (1.46 MB, 1024x1024)
1.46 MB
1.46 MB PNG
here´s some gingers for you niggas
>>
>>88773961
its 2000 times quicker to generate images
>>
I don't know how to turn off my coom-prompt brain whenever i prompt, pic not related
>>
>>88773961
Does it mean there's some small pixelated hope for cpufags like me?
>>
File: download (71).png (450 KB, 448x640)
450 KB
450 KB PNG
In case anybody was wondering, I found a better and simpler way to get the high angle shots—turns out all you really need are a lot more parenthesis. Usually the perspective on the floor is a bit off and there's room to tweak, but the following works pretty well:
>(((((high angle portrait from above)))))), closeup
"High angle portrait" does basically nothing on it's own even with the parenthesis, though "portrait from above" seems to work pretty well.
>>
getting great mechanics.
>>
>>88773979
she looks like the pores on my nose
>>
File: 40.png (465 KB, 512x512)
465 KB
465 KB PNG
how to get better guns?
>>
File: file.png (256 KB, 512x512)
256 KB
256 KB PNG
check these dubs
>>
Is this worth getting into if I just want hyper titted hijabis? How big can the booba get?
>>
File: finally.png (11 KB, 831x102)
11 KB
11 KB PNG
Sorry for wasting an image, but I'm so fucking happy this worked, finally I can switch models on colab.

Also, missed the captcha 3 times, if I got it to work, anyone can.

Thanks for reading my blog.
>>
>>88773846
These look really good. If it was larger I wouldn't mind hooking these on the wall.
>>
>>88773984
can't say for sure but its 2000 times quicker then i would say yes?
>>
>>88773884
haha thanks
saved
I'll be in my bunk
>>
>>88774006
check out >>>/h/stable
>>
File: medievil_girls3.png (2.46 MB, 1024x1536)
2.46 MB
2.46 MB PNG
bros... fuck generic anime women...

i'm in love with renaissance women...
>>
so can I make this thing spit out futa?
>>
>>88774006
yes
boobs can get very big, several times bigger than head
>>
File: u27.jpg (212 KB, 921x1382)
212 KB
212 KB JPG
>>88773894
nice
>>
File: can i see your manager.png (1.19 MB, 2048x2048)
1.19 MB
1.19 MB PNG
>>
File: unknown (1).png (723 KB, 512x512)
723 KB
723 KB PNG
>>88773996
The AI hasn't figured out how to make any item that has many different subtypes. It mixes all of the subtypes together and makes a melted mess of what looks like wires, wood, or wrapped cloth. One strategy I use to avoid this is by putting versions of the thing you don't want in the negatives (especially while inpainting). For example, you could mask the gun and put (SMG) in the positives and Rifle, Handgun, Pistol in the negatives. Should work at the correct CFG and denoising. Don't forget to add your style modifiers and artists or else you might get a photorealistic gun that fucks up the image.
>>
File: 1663910869629.png (1.21 MB, 1024x1024)
1.21 MB
1.21 MB PNG
Where's my mermaid bros at?
>>
File: mermaid1.png (288 KB, 512x512)
288 KB
288 KB PNG
>>88774043
here
>>
>>88773903
The hard part is collecting the dataset as well as getting access to 40k A100 GPUs to train it on. If I had the two I could do it myself.
>>
I've lost the will to prompt. how do I regain it
>>
>>88773979
prompt?
>>
>>88774009
go forth and proompt
>>
File: RMS.jpg (1.18 MB, 2048x2048)
1.18 MB
1.18 MB JPG
>>
File: file.png (718 KB, 512x512)
718 KB
718 KB PNG
>>
>>88774054
Realistic portrait of a woman, modern style, ginger, freckles, detailed face, beautiful face, by Adolph Menzel and Albert Lynch, digital painting, concept art, smooth, ((sharp focus illustration)), (((Soft lighting))). Sharp shadows. Intricate. High detail. ((Perfect face)) artstation hq
, Beth Conklin, Brian Froud, Edward Clark, Dave Dorman, Gustave Caillebotte, Ollie Hoff, E. William Gollings, Franz Sedlacek, George Lambourn, Elizabeth Gadd, Enki Bilal, Stephan Martiniere, Bruce Timm, Christopher Wren, Walt Kelly, Alfred Jensen, Elwood H. Smith, Cassandra Austen, Bertalan Karlovszky, Eugène Burnand, Phil Koch, Gu Kaizhi
>>
File: grid-0000.png (1.85 MB, 1024x1536)
1.85 MB
1.85 MB PNG
>>88773979
cute gingers.
>>
>>88774006
There was some guy a couple of threads ago posting some real milk trucks, like gigantic_breasts tier by danbooru standards.
>>
File: 1663624225961.png (355 KB, 448x640)
355 KB
355 KB PNG
>>88774022
Those are just anime women with a renaissance theme.
>>
>>88773832
>i found my place despite not even having lurked a single minute
bait
>>
>>88774066
>bottom right
Actually, what you are referring to as Taliban is actually Taliban/Islamic Emirate of Afghanistan
>>
File: 1663909881927.png (1.01 MB, 1024x1024)
1.01 MB
1.01 MB PNG
>>88774047
Well met
>>
File: 509438342.png (309 KB, 512x512)
309 KB
309 KB PNG
>>88774053
Don't, save yourself
>>
File: qt_v2.png (1.07 MB, 640x832)
1.07 MB
1.07 MB PNG
>>88773678
>>
File: u28.jpg (134 KB, 921x1382)
134 KB
134 KB JPG
>>88774094
interesting
>>
File: grid-0000.png (2.77 MB, 1664x1280)
2.77 MB
2.77 MB PNG
>>88774077
No, first image was just human women prompt.

this is anime touhou women in a renaissance style.
>>
>>
File: file.png (639 KB, 512x512)
639 KB
639 KB PNG
>>
What is the difference between sampling methods?
>>
File: 20.png (388 KB, 512x512)
388 KB
388 KB PNG
>>88774040
thank you
>>
>>88774107
cool background
>>
>>88773832
Nice bait but literally everyone knows that this general is nothing more but tourists from other boards and places.
>>
>>88773919
assassin
>>
please let me sleep i don't want to proompt anymore
>>
>>88774125
thx
>>
>>88774126
This. I don't even use /g/. Just here for pictures.
>>
File: u29.jpg (186 KB, 921x1382)
186 KB
186 KB JPG
>>
Test
>>
covered booba
>>
>>88774103
I was just being a bitch about the cartoony style. But it's a nice enough style and none of them have nasty Mucha hair so I guess I should just be greteful.
>>
>>88774122
You're welcome, and good luck.
>>
cyberbooba
>>
The Apple Experience

(Steve's in hell now for all his crimes against humanity!)
>>
File: 5094383411.png (351 KB, 512x512)
351 KB
351 KB PNG
>>88774050
Idk you so maybe, but there's actually a lot of know-how you need to have to do the initial training successfully.
>>
File: file.png (470 KB, 768x512)
470 KB
470 KB PNG
This is pretty good but when's the next big upgrade coming?
>>
>>88773678
is there a place to do img2img online?
>>
>>88774142
Damn. Share the full prompt? I've never gotten that kind of skindentation
>>
Is there a way to save the result every N steps without starting from 0 every time? There's an option to show it in the UI, but only the final step appears in the output dir.
>>
>>88774168
yes
it sucks
>>
>>88774135
I'm ... I'm gonna proooooooopt! Ahhhh!
>>
File: test.png (400 KB, 512x960)
400 KB
400 KB PNG
found some fixes, and not bad
>>
File: prettysmile.png (457 KB, 512x512)
457 KB
457 KB PNG
>got 3 days off
>have fat blocks of smoked gouda and a box of ritz
>gf went with her bitch whore best friend to Atlanta for the weekend
>apartment all to myself
yep, it's time to proompt
>>
>>88774010
thanks. I haven't gotten around to installing upscayl or r-esrgan yet.
I'm an ITsupportlet and getting a simple github repository installed is an entire weekend project for me.... -.-
>>
>>88774142
don't forget you can upscale your images using cupscale
>>
File: 1663947735647-1.png (1.47 MB, 1024x1024)
1.47 MB
1.47 MB PNG
ginger lolis
>>
How do you go about finding good artist combinations? Just rolling until you get something works?
>>
File: file.png (284 KB, 512x512)
284 KB
284 KB PNG
it doesn't understand me
>>
>>88774207
yes.
>>
>>88773832
Anon, Activate your Windows!
>>
>3 paperspace instances open for maximum proompting
Might get drunk and open more.
>>
File: download (14).png (468 KB, 512x512)
468 KB
468 KB PNG
>>88774207
Think about what you are trying to make and find the artist that does that the best in your opinion. Go learn lots of artists.
>>
>>88774171
>>88774171
> realistic proportions, highly detailed, (((skindentation))), (((PERFECT_FACE))), (((lvoluptuous_breasts))), skinny waist, wide hips, tight_clothes,

These are the ones from the filename. Experiment with it.

>>88774203
Hey thanks, I will upscale my favorites!
>>
>>88774135
>>88774135
I'm ... I'm gonna proooooooompt! Ahhhh!

I can has promt please?
>>
File: 1456557199.png (281 KB, 512x512)
281 KB
281 KB PNG
>>88774065
Thanks, have this tiling frog stained glass in the style of Murakami.
>>
>>88774231
I was wondering about the artists/model because I already read the filename.
>>
>>88773930
If you care, you can always set a power limit in Afterburner.
>>
File: file.png (617 KB, 768x512)
617 KB
617 KB PNG
It's over for artists
>>
>>88774175
>https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts
If you mean by saving steps of the sampling process then scroll down to the bottom script.
>>
>>88774247
Ah, I'm sorry, I made it yesterday so I don't remember exactly. Don't think I used an artist however. Sorry again! I think they are all renaissance paintings however as that is the folder they are in
>>
File: cuteCUTE.png (293 KB, 512x512)
293 KB
293 KB PNG
Cute.
>>
File: 1658389313665.png (332 KB, 512x512)
332 KB
332 KB PNG
>>
man it sucks waiting 20 minutes for a prompt to finish only to get bad results.
>>
>>88774257
>free porn available to anyone all the time
>people still spend thousands of dollars on semi nude photographs of women on onlyfans
I don't think so, Tim.
>>
>>88774264
is that the one guy was making webms out of?
>>
File: 3974439534.png (309 KB, 512x512)
309 KB
309 KB PNG
>>88774230
looks like a Cleanrot Knight
>>
>>88774207
My process was to find a style/school of art that I liked, then review each artist and tote the specifics of what I liked about them. For example, I like to generate buildings, portraits, and landscapes. So I rated each artist on how well they did each of those. Make sure to check that they're in the data, too. I also took notes on whether they had detailed backgrounds, whether they painted both men and women, if the people they painted were all very samey, if their landscapes included a lot of people, etc. Then you can make an educated guess on what will work well for your purpose, but be prepared to find that your assumptions were wrong once you start making x/y grids comparing results.
>>
>>88774278
how do you think artists feel
>>
>>88773754
prompt??
>>
>>88773965
>>88773952
This is result of official
https://pytorch.org/get-started/locally/
stable, windows, conda, cuda 11.3
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
>conda list -f pytorch
1.12.1 py3.9_cuda11.3_cudnn8_0
>python -c "import torch; print(torch.cuda.is_available())"
false
>>88773878
here is result of installing requirements from Voldy txt files
pip install -r requirements_versions.txt
pip install -r requirements.txt
>pip list
torch 1.121+cu113
>python -c "import torch; print(torch.cuda.is_available())"
false

All of this is really annoying since both versions of installation worked just fine few days ago but now it just shits itself.
Could someone else show me what results they are getting from their versions of pytorch?
>conda list -f pytorch
>python -c "import torch; print(torch.cuda.is_available())"
>python -c "import torch; print(torch.version.cuda)"
.python -c "import torch; print(torch.zeros(1).cuda())"
>>
>>
>>88774278
>20min
what the fuck, why dont you colab, you cretin
>>
>>88774271
No worries.

If you're using the webui version You can drag your .png files into the png.info tab to get your full prompt for any picture. Really handy when you're like me and forget the prompt you wrote a day ago. I'll start experimenting now
>>
>>88774291
same as last thread >>88772620
>>
>>88774278
And I feel impatient when it takes 20 seconds to load a landscape
>>
>>88774294
what gpu do you have?
>>
File: file.png (574 KB, 768x512)
574 KB
574 KB PNG
How do I upscale a result I like?
>>
>>88774257
nice house.
>>
File: 1663948199025-1.png (1.4 MB, 1024x1024)
1.4 MB
1.4 MB PNG
you can never stop me waifu-posting
>>
>>88774120
everyone uses euler-a at 20-ish steps
why use anything else
>>
When I try to inpaint with webui (via the FOSS Krita plugin) it tends to ignore my prompt and erase the area. I want to do something like inpaint a face and have it generate just a new face. Is that possible?
>>
File: yum.png (448 KB, 512x512)
448 KB
448 KB PNG
>>88774281
It's the parasocial aspect they are paying for usually. Just need to combine the story telling AI with the image one and you can replace them mostly.
>>
>>88774286
>ariel
>white
brooooooo...
>>
>>88774303
because I've heard horror stories about it generating very questionable stuff despite peoples prompts not containing anything related to that.
>>
File: uinput.png (797 KB, 921x1382)
797 KB
797 KB PNG
>>88774234
Glad you like it. It's img2img, trinart2.

(full body), (3d), [(city lights):.05], (((realistic eyes))), (starry eyes), ((unicorn horn on head)), (panties), underwear, ((cute face)), beautiful face, sexual, (intricate), ((sexy)), tone mapped, ambient lighting, highly detailed, digital painting, [(sailor moon):.1], anime, ((very cute and smug manga girl)), sharp focus, hime cut, twin tails, (neon cyberpunk), [lingerie]
Negative prompt: out of frame, extra fingers, (((mutated hands and fingers))), ((poorly drawn hands)), ((poorly drawn face)), (((mutation))), (((deformed face))), ((ugly)), blurry, ((bad anatomy)), (((bad proportions))), (((extra limbs))), extra face, breast, ((double head)), ((extra head)), ((two girls)), (big breast), (((extra feet))), penis, ((two horns)), ((extra horns)), (((boy))), male, (((ugly eyes))), (((imperfect eyes)))
Steps: 39, Sampler: Euler a, CFG scale: 10, Seed: 824457099, Size: 512x768, Batch size: 8, Batch pos: 0, Seed resize from: 512x512, Denoising strength: 0.83, Mask blur: 4


Input file attached.
>>
>>88774294
IF
>torch.cuda.is_available()
is false then its a driver or environment or something problem. assuming this isn't a vinoboy i can't deal with their bullshit.

if you didn't use conda you're a cunt because you can easily see when that's on or off before launching. it really should be initialized by default but i dont think anyone will.

anyone know if there's an equivalent
>torch.cuda.is_available()
for whatever AMD uses?
>>
File: download (8).png (1.5 MB, 1536x2304)
1.5 MB
1.5 MB PNG
Anyone have any hints for better inpainting? My eyes always seem to look like shit if I don't get them right by the prompt.
>>
>>88774326
Do not let jannies or trannies take away your AI wife.
>>
>>88774337
and whats the issue with that. you fear the feds? a true prompter knows no fear.
>>
File: Pearls_n_Nymphs.jpg (586 KB, 1534x1533)
586 KB
586 KB JPG
>>88774043
I added Mermaids and Nymphs to my proompt and this what I got....
>>
>>88774284
>https://pastebin.com/h3qXa8hZ
Nah though you could make a webm out of the output but it would only be of noise converging into the image upon reaching the final step. The above script is from an anon on /h/ that makes webm.
>>
>>88774357
these look amazing anon, mind giving prompt?
>>
>>88774320
>>88773867 gtx 1080 (nvidia-smi info: driver 473.62 CUDA 11.4), like I've said before, I was able to use Voldy repo beforehand but for whatever reason it all stop working today.
>>
>all you need to build a magic art machine is calculus and linear algebra that was already worked out in the 1800s
Isn’t that kind of bonkers when you think about it?
>>
>>88774345
do you have restore faces on
>>
It's so piss easy to create character art for any indie game now
It gets a major money sink out of the way immediately
Fucking incredible
>>
>>88774373
>all you need to do is multiplication and addition that was figured out in 1650
crazy to think about it doesn't it
>>
>>88774373
oh i hate this
>>
>>
>>88774382
it's very inconsistent but significantly MORE consistent than a fucking artist.
>>
File: u31.jpg (206 KB, 921x1382)
206 KB
206 KB JPG
>>88774342
>>
>>88774382
Ok now animate them.
Wait let me guess:
>just give it another year or two, this technology is progressing incredibly quickly
Did I get that right?
>>
>>88774094
but what am I going to do without prompting? I'll have to play League of Legends again or something. I don't want to go back
>>
>>88773894
>>
>>88774343
>driver or environment or something problem
Like I've said before, the usual tutorial that worked for me was create the new conda environment without setting up version of python since running the pytorch official 'conda install pytorch' would install python version that worked just fine on my system.
I have not fucked around my main cuda or nvidia driver, so I am running out of ideas on what could cause this.
>>
File: 0009.jpg (175 KB, 510x514)
175 KB
175 KB JPG
so many waifu. but are husbands allowed here?
>>
>>88774412
yes :)
>>
File: index.png (277 KB, 512x512)
277 KB
277 KB PNG
I swear you guys are all trolling me, this is fucking shit
>>
>>88774425
well whether you have fucked around with it or not it is not there. your driver is gone buddy. try reinstalling official cuda whatevers.
>>
>>88774370
Maybe you installed it as admin?
>>
>>88773914
Click the Hires fix. box
>>
>>88774437
If you don't post your prompt nobody can really help you
>>
>>88774354
I imagine you live in the middle east or the deepest jungle of brazil.
>>
>>88774437
try 50 steps and 16 cfg for anime girls
>>
Having some problems with openvino.
Can someone help, please?

Exception in Tkinter callback
Traceback (most recent call last):
File "/home/ai/miniconda3/envs/vin/lib/python3.9/tkinter/__init__.py", line 1892, in __call__
return self.func(*args)
File "/home/ai/stable_diffusion.openvino/pygui.py", line 147, in restore_to_queue
self.queue_items.append(QueueItem(self.text,inp.input, self))
File "/home/ai/stable_diffusion.openvino/pygui.py", line 369, in __init__
self.data = QueueData(data_list, force_number)
File "/home/ai/stable_diffusion.openvino/pygui.py", line 457, in __init__
if not self.upscale:
AttributeError: 'QueueData' object has no attribute 'upscale'


Followed the CPUTARD guide on the OP and set all of the settings correctly. For "upscaler" in settings, I set
/home/ai/stable_diffusion.openvino/realesrgan-ncnn-vulkan-20220424-ubuntu/realesrgan-ncnn-vulkan

but the error above pops up in the terminal when I click "Generate".

The model that is supposed to be downloaded in my .cache is not being downloaded.
>>
File: greg rutkowski.png (909 KB, 1536x640)
909 KB
909 KB PNG
>>88774205
>>
>>88774437

Finally, you caught us, we all learned to draw and shit, what a healthy prank, right? Hahaha
>>
>>88774271
What model is? WD?
>>
>>88774427
he's way less downsy than the last one i saw, very nice.
>>
>>88774368
this is on Basujindal OptimizedSD

"ominous portentious sky, mermaids, nymphs, blood red clouds, huge ragged pirate galleon made of barnacles, background lightning and rain of blood and morrslieb warpstone moon made of skulls, william turner, bioluminescent details, god rays, abyssal sketch, as imaged by lovecraft, in frame" --H 512 --W 768 --seed 465 --n_iter 7 --n_samples 2 --ddim_steps 28 --sampler euler_a --scale 13
>>
>>88773920
When I add the prompt
>anatomically correct
I get a lot of stuff like this. I made the mistake of adding it to coom prompts which is where I learned it
>>
File: 1643226628862.png (521 KB, 512x768)
521 KB
521 KB PNG
>ask for a "tender smile"
>get some NTR face
ai please
>>
File: file.png (576 KB, 512x512)
576 KB
576 KB PNG
>>
>>88774471
BUILT
>>
>>88774447
Yeah I figured out a few different things, that is what I get for being stupid and not reading.
Thanks again though
>>
>>88774471
someone did up smile/emotion charts soemwhere
>>
>>88773933
get fuckt lmao I’m gonna make it 6
>>
>>88774467
thanks
>>
>>
I get
RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'
what should i do?
>>
>>88774437
skill issue
>>
>type "goatse" only in the prompt
>gets various imagery focusing on a circle
it knows doesn't it
>>
>>88774488
Literally a random elden ring location.
>>
>>88773124
Please post this to /h/ or something. And please keep them cooooming.
>>
are there options for CFG from the base build?
I don't see it outlined in the github, but I see it in the webapp wrapper some people are using
>>
>>88774461
stop spoonfeeding!
>>
>>88774471
You just know.
>>
>>88774278
unless you are using the cpu you should just be using DDIM at 20 samples in big batches to know if your prompt needs tweaking. That takes seconds, some minutes at most. Not 20 minutes certainly
>>
>>88774471
looks like that princess from Overlord
>>
File: file.png (635 KB, 512x512)
635 KB
635 KB PNG
>>
File: rv.png (1.24 MB, 1024x1024)
1.24 MB
1.24 MB PNG
>>
>>88774466
downsier are gentler, you dont want a bad uncaring smart guy, body has too much of everything, has to be used with love.
>>
>>88774429
If this thing was already 100% perfect at illustration then I would believe you that animation is just around the corner.
But it still has glaring flaws even when it just comes to static illustrations, and those problems will just get amplified and become more pronounced when it tries to do animation.
It has a tendency to add in extra limbs, blend things together in weird ways, etc. because it doesn’t actually “understand” what anything is or how things interact in 3D space, it just knows how to get a close enough approximation to what something looks like.
“Close enough” is good enough for a single illustration (sometimes) but carried over many frames of animation that are supposed to show things moving and interacting with believable motion and weight and the problems will just get amplified.
>>
>>88774373
damn i wish i paid more attention to calculus in school
>>
File: 318083476.png (299 KB, 512x512)
299 KB
299 KB PNG
>>88774335
Oh shit, I'm sorry
>>
>>88774530
Prompt for making more of these disgusted-looking women? The look of mild disdain you managed is very nice.
>>
File: u32.jpg (234 KB, 921x1382)
234 KB
234 KB JPG
>>
Is 2080 ti the best GPU under $500 for stable diffusion?
>>
File: u33.jpg (183 KB, 921x1382)
183 KB
183 KB JPG
>>88774564
i need to go home and stop prompting
>>
>>88774557
much better!
You know perfectly well that whitewashing is racist and not allowed under global rule 3.b!
>>
>>88774492
change runtime type from cpu to gpu (assuming you're on colab)
>>
File: 1663941524651659.jpg (573 KB, 1726x814)
573 KB
573 KB JPG
>>88773850
lots of trial and error, like lots and lots and lots.
Each source image needs a different level of denoising in the 3-5 range, doing small increments till you find something that does not obliterate the image but also starts producing interesting results then generate 100s at that and wait for something good to pop out.
Pic related, it needs posting whenever someone asks this question.
>>
>>
>>88774569
I found a 3060 (12GB vram) for 500$. 400$ because I bought used for sloppy seconds and free return if it shits the bed.
>>
File: 00002.png (1.75 MB, 2048x2048)
1.75 MB
1.75 MB PNG
>>
I think adding Terese Nielsen to my artist list really cleaned up my body proportions and is curtailing some mutations.
>>
>>88774142
>realistic proportions
>>
File: 3473021153.png (264 KB, 512x512)
264 KB
264 KB PNG
Honey wake up new UMP just dropped
>>
File: u36.jpg (237 KB, 921x1382)
237 KB
237 KB JPG
i'm gonna unicornnnnn
>>
File: grid-0001.png (2.32 MB, 2048x1024)
2.32 MB
2.32 MB PNG
cute toys
>>
>>88774569
MSI is selling refurbished 12gb 3060 for $329 on ebay. If you want to risk it
>>
File: 1609768263971.jpg (52 KB, 646x595)
52 KB
52 KB JPG
>>88774074
REAL honkamaloojas
>>
>>88774629
I like this artist list. could you share?
>>
File: tj.png (1.23 MB, 1024x1024)
1.23 MB
1.23 MB PNG
>>88774563
Nothing specific, probably got lucky with the seed.

6384370359831525203

"hyper realistic render, female, young, tight clothes, cleavage, soft lighting, sharp focus"

30 steps. 12.5 scale. CodeFormer + Ee 0.4 to fix face a bit.
>>
Which sampling method do you use?
>>
not so hot on some of the color choices but using runway model and fashion model in the prompt to get it to hide hands works nice
>>
>>88774642
i wonder if it can do girls frontline characters or it just gets really confused
>>
>>88773832
this cant be real
>>
>>88774658
by ((((Yoji Shinkawa)))), ((Terese Nielsen)), (Carne Griffiths), (Yoshikazu Yasuhiko), ((Wadim Kashin)),
>>
>>88774661
Euler-A
>>
>>
File: tomb8.png (1.37 MB, 1024x1024)
1.37 MB
1.37 MB PNG
booba
>>
>>88774661

Euler_a at low steps to search for seed, then Euler at high steps to generate the image I'll actually use.
>>
>>88774659
Suppose that's how is, a good reminder that massive double-layered prompts aren't the only way to getting what you want. Sometimes setting a vague baseline that aims for what you want and letting it fly is just as good. Thanks for it.
>>
>>88774707
More steps equals better image right? But what's the difference between euler a and euler?
>>
File: 1663949481530817.png (517 KB, 512x448)
517 KB
517 KB PNG
how do i make something like this?
>>
File: u37.jpg (260 KB, 921x1382)
260 KB
260 KB JPG
>>88774644
>>
File: 1663949255088-noauth.jpg (30 KB, 512x512)
30 KB
30 KB JPG
How do I get a much more detailed output? I'm just using the img2img site.
>>
>>88774661
>>88774659
For this? Using NOP on Colab I use the Default Scheduler which I think is pndm. Precision is autocast. Model ID is SD1.4. Diffuser version is latest.
>>
If I download the waifu diffusion model, can I swap between it and the regular model freely, or will I need to replace the model file in my folder each time?
>>
>>88774424
Very nice, prompt?
>>
you stack artists like Rutkowski to make your images "better", I stack Roger Dean and gemstones, we are not the same
>>
>>88774733
put in models folder, have a good read at everything you can do in the settings tab
>>
File: grid-0002.png (2.53 MB, 2048x1152)
2.53 MB
2.53 MB PNG
apparently my 3080 can only render 1024x576
>>
File: u38.jpg (205 KB, 921x1382)
205 KB
205 KB JPG
>>88774725
>>
>>88774740
>>
>>88774723

The "_a" samplers add noise at every step, while the others just refine what they have over and over.
>>
>>88774723
An A
>>
File: grid-0003.png (2.38 MB, 2048x1152)
2.38 MB
2.38 MB PNG
Toy Story!
>>
File: 21213index.png (198 KB, 512x512)
198 KB
198 KB PNG
New fetish unlocked
>>
>>88774231
>(((PERFECT_FACE)))
>cuts it entirely
kek
>>
>>88774627
loooool
you just made me google the name of that ancient game I had dim recollections of gaming as a very young boy...

fucking Missile Command

>>88774629
these are amazing
>>
Should I buy the upcoming RTX6000 for SD? I can afford it if I make some sacrifices.
>>
>>88774724
go to the store and get all these things
>2 carrots
>2 limes
>2 large black olives
>a lemon
>a cherry tomato (slightly unripe)
>a courgette
>a bag of kale
then I’d just go from the reference image and arrange on a plate as necessary
>>
File: 10645-uhhoutput.png (574 KB, 512x704)
574 KB
574 KB PNG
Can anyone clue me in on what the prompts of the form [female:male:0.4] mean?
I have a... project, which may benefit if the syntax works the way that I think it works
>>
>>88774768
Looks shitty but it's got potential
>>
>>88774456
Bump!
>>
>>88774777
trips for truth
>>
File: AAHHH.png (466 KB, 512x512)
466 KB
466 KB PNG
I haven't fucked around with portrait art much so this might not be helpful but when I was messing around with Jason Edmiston I got a lot of extremely expressive faces. Here's the prompts:

(((Jason Edmiston))),(highly detailed), high res, (((extremely detailed face))), ((dramatic lighting)), one person, (close up of woman's face), (((angry expression))), mad face, wrinkles
>>
>>88774787
Not clear yet. One of the things that may happen soon is a possible reduction in ram usage/increase in speed:
https://github.com/TheLastBen/fast-stable-diffusion

But then bigger models will also happen and stuff and things. Where the actually capable weeb furry porn freedom data settles is not easily predictable.
>>
>>88774803
lol i screwed up my jason edmiston notes so i have no idea what he's supposed to influence. it was having fuck all on my distant characters in landscapey thing. will have to re-check eventually.

>>88774804
eh
>>
very nice
>>
>>88773924
> Thomas the Tank
is that rhyming slang?
>>
>>88774778
thank you brother.
>>
File: 1663950100447-1.png (1.35 MB, 1024x1024)
1.35 MB
1.35 MB PNG
I peak inside an artist's studio
>>
>>88774769
>the perfect woman doesnt exis-
>>
>>
>>88774827
cozy.
>>
>>88774794
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#negative-prompt
>>
File: u39.jpg (296 KB, 921x1382)
296 KB
296 KB JPG
>>88774758
ok i'm finally going to go home
>>
>>
You guys capped the images, so here we go...

>>88774851

>>88774851

>>88774851
>>
>>88774698
Classy
>>
>>88774796
>>88774456
Turns out my manual value of "x2" for upscale was not accepted (but is what I used with the esrgan executable itself), so I had to choose x4 via the GUI. I suck at GUIs.
>>
Quite nice.
>>88774698
>>
>>88774698
proompt? they look soft.
>>
>>88774881
(Frank Cho), [Pieter Claesz], [Alphonse Mucha]. Dry acrylic painting. (((2girls))). ((Stacked Erotic Secretary, Bryce Dallas Howard), (beautiful Redhead)) . ((erotic slime office pantsuit)). (choker). (detailed realistic faces), sensual eyes, ((erotic aspect)), ((sensual lips)), high detail, intricate, (collarbones).
>>
>>88774899
thank you, anon :)
>>
>>88774525
I tried at a batch of 15 and it took 20 minutes to do 10
>>
>>88774724
see
>>88774619
>>
>>88774842
I assume by home you mean prison
>>
>>88774840
I actually meant (what I now know is referred to as) Prompt Editing, but that was close enough, ty m9
>>
I've been using google collab in OP and it usually works for 2-3 generates before images stop coming back or don't generate at all and nothing shows up in the UI.

I thought it might be a storage issue but it appears the disk is not full so I'm not sure what causes it. Could it be the disk space?
>>
>>88774971
Oh, that is actually my bad. I meant to link you to Prompt Editing.
>>
>>88774603
its a native install for amd
>>
To run it offline, all I need to do is keep the cache files (that it downloads) right? Or is there more to it?
>>
>>88775097
that all depends on what "it" is anon, what is it?
>>
>>88775120
My bad, I forget you're not inside of my head. SD and its forks.
All that I've used so far, upon first use, download files to .cache.
>>
>>88775152
Yes
>>
>>88775152
sd-webui? the conda build downloads a bunch of cache stuff to /tmp/, not sure if elsewhere. some python venv might download to .cache?

either way the cache stuff shouuuuld be killable if you've got the packages in your environment's site-packages...? that's like the library-path for pythons.
>>
>>88775194
>sd-webui?
In my case, basujindal's optmized SD, Voldy/AUTOMATIC's and openvino.

I'll have a look at /tmp, altough I hope it doesn't put anything internet-related there.
I keep my .cache as tmpfs, so I'll try populating it after a reboot and see if it tries to download other stuff.
I'm not sure about the site-packages, haven't noticed any activity there after the first setup.

I hope >>88775178 is right as just repopulating my cache would be easy.
>>
>>88775259
the guides are very nonstandard for some reason and ultimately quite shit. if you use conda you're going to need to be super consistent about activating it or you won't be in the environment with shit in your path.

if you don't use conda then enjoy things being shat all around your filesystem willy nilly like .cache. this is expected behavior.

repopulating .cache should always be easy, it just needs to actually exist.
>>
File: Nightgira-Fuel.png (172 KB, 270x270)
172 KB
172 KB PNG
>>
So how intense is the image generation process? Is it gonna scald my 2060 or is this shit relatively tame for modern cards?
>>
>>88776419
it's always going to be 100% usage while you're generating, gpu power just makes it go faster. and 2060 is pretty weak



Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.