Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
Pretty, yet creepy.

1690191633330768.jpg
 
  • Like
  • Thunk-Provoking
  • Winner
Reactions: Vash the Stampede, Xenomorphs Are Cute, Irrational Exuberance and 11 others
Catgirl Tyranid said:
We point fingers over who's to blame or we can actually do what's within our power to roll back the damage unlimited porn has and will cause. Quit consuming porn and contributing to that market so these leeches stop making money, tell anyone who asks that porn is gross regardless of how it's made, and carefully teach your kids what a healthy relationship is and how to strive for it.

And to make sure this is actually on topic, giving boys access to unlimited AI porn right as they reach puberty will collapse civilization. Don't give your kids smart phones.
Click to expand...
He has a point tho, for all the white knighting the fact is most women in porn choose that because its easy money, specially now in the camwhore era when you don't even have to do sex scenes and can get away with just some nudity as long as you get some loyal paypigs.

Also:r28ybdp8p7l11.jpg

I'll argue that far more men have been ruined by anime than by porn. Like the other guy said few men are deranged enough to be open about their porn use but look at weebs and tell me how many are ashamed of that.
Puff said:
Just fine. At 512 square I get 1 to 2 iterations a second. Make sure to install xformers.
Click to expand...
How long it takes to do a 2048 square?
Gourmet Race said:
Thoughts?
Click to expand...
Well its clear adobe its shitting themselves because AI is going to destroy their bottom line.
Gourmet Race said:
Hopefully, artists see through this.
Click to expand...
They wont, most are malding about this and demanding luddite-levels of technological restrictions.
Omen_Sulk said:
Pretty, yet creepy.

View attachment 5230401
Click to expand...
Not really creepy, more like uncanny valley-ish. The girls look less real asian and more "korean plastic surgery overload".
 
  • Like
Reactions: Catgirl Tyranid
  • Informative
Reactions: Ether Being and Mr.Miyagi
The Mass Shooter Ron Soye said:
Click to expand...
Heard a little about this, but didn't realize it was already available. It's interesting. Seems like it is an improvement on the old base, though I'm not experienced enough with it all to really know the finer details about performance. Not really sold on it being much better with hands like the blog says, but that might just be my luck. If anybody wants to give it a quick shot, you can use the clipdrop site mentioned in the blog. You get a handful of gens before they ask you to make a free account/sign in with whatever, and there's a queue, but it's a pretty simple way to give it a look.
stable-diffusion-xl(1).jpgstable-diffusion-xl(2).jpgstable-diffusion-xl(3).jpgstable-diffusion-xl(6).jpg
 
  • Like
Reactions: Ether Being and The Mass Shooter Ron Soye
cybertoaster said:
How long it takes to do a 2048 square?
Click to expand...
There isn't a viable model yet to do 2048 text2pic with the accuracy of stable diffusion 1/2/XL. SDXL just leaked and it was trained on 1024. Even if you do the upscaling pic2pic hack you are going to run into issues with VRAM unless you are running a 4080(Cutting it super close and need it headless) 3090/4090 or a Tesla p40.
 
  • Agree
  • Informative
Reactions: cybertoaster and std::string
Deadwaste said:
impressive, but i still dont know how youre actually supposed to download these things and make them work. nor how to get into other stable diffusion shit like controlnet or whatever else there is i can use with a111
Click to expand...
There's a link to a github with step by step instructions for getting the environment set up, which also links to huggingface to get the actual model. huggingface has their own internal git-like download page for each model, where the actual model is denoted by a red box icon. You have to have an account to download it. It's a bit convoluted.
 
macrodegenerate said:
There isn't a viable model yet to do 2048 text2pic with the accuracy of stable diffusion 1/2/XL. SDXL just leaked and it was trained on 1024. Even if you do the upscaling pic2pic hack you are going to run into issues with VRAM unless you are running a 4080(Cutting it super close and need it headless) 3090/4090 or a Tesla p40.
Click to expand...
I've been throwing things at a H100 and yeah it just makes models eat shrooms faster. You can stretch it to about 1.3x what the model was trained on before it just collapses.
 
I played around with 1.5 after a while of absence. Depth maps are new for me and they're pretty powerful. you can just take some random picture off the internet like e.g.:
car.jpg
and then use a depth map (with weak controlnet influence) and prompting to make something entirely else but similar out of it, like e.g.:
car2.jpg

Then you could improve on it with inpainting and photoshop and nobody will ever find out you're not a real artist

You can also feed generations back into it to generate depth maps out of those generations etc. etc. lots of possibility to actually affect the output. The nice thing is that you can completely bypass the inherent SD incoherency in objects by helping it along the way.
 
  • Like
  • Informative
Reactions: indomitable snowman, Post Reply, Xenomorphs Are Cute and 4 others
I don't really use ControlNet all that much, but when I do it's usually OpenPose and Reference.
 
macrodegenerate said:
There isn't a viable model yet to do 2048 text2pic with the accuracy of stable diffusion 1/2/XL. SDXL just leaked and it was trained on 1024. Even if you do the upscaling pic2pic hack you are going to run into issues with VRAM unless you are running a 4080(Cutting it super close and need it headless) 3090/4090 or a Tesla p40.
Click to expand...
Interesting, but what about those paid AI sites that offer that resolution for premium subs? I seen those posted on reddit and were that resolution, what are they using to do it?
 
cybertoaster said:
Interesting, but what about those paid AI sites that offer that resolution for premium subs? I seen those posted on reddit and were that resolution, what are they using to do it?
Click to expand...
They're probably upscaling a 512x512 with Real-ESRGAN or Remacri or something. I don't have any issues with getting a 2048 or 3000px image through upscaling, but yeah, no way I can get a raw generation that size with txt2img.
 
  • Agree
  • Thunk-Provoking
Reactions: std::string, woowie queen, Baraadmirer and 2 others
cybertoaster said:
Also:

How long it takes to do a 2048 square?
Click to expand...
About 24 sec per iteration. Basically, to make big images, I get the small image the way I want, upscale and then image to image to clean it up.
 
So, these were made by SDXL with a pixel art lora.

pixel2.jpgpixel3.jpgpixel4.jpgpixel5.jpgpixel6.jpgpixel7.jpgpixel8.jpgpixel9.jpgpixel10.jpgpixel1.jpg

Since pixel art is really lossy re: fidelity/information, it has the advantage of not really pointing out mistakes. I didn't bother editing them, but it would've been easy to. All I did was run a script for limiting the color palette and dithering, if necessary. It wasn't even really good and could've been done better. (If they're blurry, it's because your browser is scaling them, try opening them in a new tab) I also accidentally everything to jpg, sorry.
 
  • Like
  • Winner
  • Feels
Reactions: SaidTheSpider, Vash the Stampede, Seymour Glass and 11 others
Mr.Miyagi said:
They're probably upscaling a 512x512 with Real-ESRGAN or Remacri or something. I don't have any issues with getting a 2048 or 3000px image through upscaling, but yeah, no way I can get a raw generation that size with txt2img.
Click to expand...
IIRC Leonardo uses the 768 and 1024 models but yeah it's just upscaling.

Creating multi-monitor desktop backgrounds (like 5920x1080) is some voodoo I haven't managed to unlock yet, though I've seen people pull it off.
 
You must log in or register to reply here.
Back