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

Probably An Autist said:
What GPUs are you guys using? My 2060 only has 6GB of VRAM and is kinda low on CUDA cores so it struggles on image size.
Click to expand...
I'm on a 1060 6gb using the --medvram command. The website does mention going over 512x512 tends to produce more errors, so I've been doing a batch of 4 then upscaling the ones I like. I find 512x1024 works fine for landscapes but it tends to break anything else pretty quickly.


(mentions 768x768 being fine for novelAI but I've seen more errors doing that too)
 
  • Informative
  • Agree
Reactions: Shiawase and A Hot Potato
Argrabius said:
The only people losing against this are people whose business model was “I am the only one who makes your flavor of depravity, pay up or no fap.” I think we could see a renaissance of digital art since artist comissions will be based more on true high skill and personally developed art styles since if you just want fringe material ai can do that
Click to expand...
The idea that there will remain some upper crust of true professionals is delusional.
 
  • Agree
  • Autistic
  • Mad at the Internet
Reactions: Toolbox, umami's milk, Shiawase and 2 others
Probably An Autist said:
What GPUs are you guys using? My 2060 only has 6GB of VRAM and is kinda low on CUDA cores so it struggles on image size.
Click to expand...
My gaming rig with a 2080ti or my laptop with a 1050. I tend to stay on the laptop primarily when I proompt since I don't want my electric bill to go through the roof but when I find a good prompt with a decent success rate of what I want I move to the gaming rig to mass produce a bunch of shit.

raytwace.png
 
  • Like
  • Informative
  • Feels
Reactions: a feel, Cowboy Kim, Probably An Autist and 7 others
TheGuntinator said:
My poor AI dungeon. AI tech is so interesting, but some people just feed it garbage prompts trying to jerk off and ruin everyone else's experience. The fucking nightmares that AI came up for me when i just wanted to slay some orcs and raise the dead. I was dragged kicking and screaming to The Bone Zone.
Click to expand...
My understanding (though it's been a while since I read the papers) is that an important aspect of AI Dungeon/Novel AI/Stable Diffusion is that they use pre-trained weights. So unlike something like Tay, designed to learn and change through user interaction, these models won't be affected by user input. So no matter how many people put in text about fucking or generate explicit pictures, it won't "corrupt" the outputs you receive. The main reason why the models steer so easily into explicit material is just a byproduct of the training corpus used, like Common Crawl. AI Dungeon also suffered from fine-tuning by Latitude on freely available Choose Your Own Adventure stories containing some questionable material (and quality), while NAI's SD model, with such a huge volume of specific material from Danbooru, naturally gravitates to anime women with huge titties.

Since training a wholly original model is both expensive and time consuming, our best bet is just using blacklisting functions for things we don't want, along with fine-tuning the models on specific things we want it to do better at, creating niche model versions for certain topics.
 
  • Informative
Reactions: Cowboy Kim, A Hot Potato, TheGuntinator and 2 others
Deluxe2 said:
You need to use Danbooru tags demarcated with commas to use NovelAI effectively. Go to any SD general on 4chan for examples of how tagging should work

This is a general template for how tag order should be arranged for solo focus images (like porno)
Click to expand...
Dare I ask what a "cum prompt" entails?
Inspector Rex said:
Can someone try to generate some B/W clip art? Can it draw clean simple lines?

At some point someone is going to train this thing to draw UIs, and that's when graphical designers should start worrying.
Click to expand...
I get this with a clipart, black and white, best quality, masterpiece prompt:
1666209613473.png

It's probably easier to just make a line in Paint or Photoshop. I got this when trying to specify a single horizontal line:
1666209742164.png
Colon capital V said:
now put in just "faggot" and see what it outputs
Click to expand...
I went through a few iterations and got this:
1666209798557.png
 
  • Like
  • Informative
Reactions: TheGuntinator, Inspector Rex, A Hot Potato and 2 others
cerulean hydra said:
I'm on a 1060 6gb using the --medvram command. The website does mention going over 512x512 tends to produce more errors, so I've been doing a batch of 4 then upscaling the ones I like. I find 512x1024 works fine for landscapes but it tends to break anything else pretty quickly.


(mentions 768x768 being fine for novelAI but I've seen more errors doing that too)
Click to expand...
I am on a 1060 6GiB and found medvram wasn't needed anymore using the automatic webui and with xformers my generations are around 25s at 512x768. The reason errors mount at anything above 512x512 is all the models are trained at that.
I'm currently using this:
Code: 
set COMMANDLINE_ARGS=--listen --xformers --deepdanbooru
 
You must log in or register to reply here.
Back