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

New wojaks dropping

yeschad.pngyeschadHD.jpg
tradwife.jpgtradwifeHD.jpg
gothgirl.jpggothgirlHD.jpg
doomer.jpgdoomerHD.jpg
timemachine.jpg
(as always, click the thumbnails for high res. there is a lot of detail in these that the thumbnails don't show.)

These were more complicated to do. I had to go back and forth between photoshop and stable diffusion to fix errors between re-images. Sometimes stable diffusion simply would not recognize things, such as Doomer's cigarette. I've come to view stable diffusion as less of a replacement to artists and more of another tool to use when making art. Additional touchups are almost mandatory if you want something to look a certain way.
 
Last edited:
Last edited:
Holy fuck mate those are fucking sick, if you don't mind me asking how did you manage this???
I'm using Midjourney. Prompt is:
DVD screengrab from a James Cameron fantasy sci-fi film 1988 --ar 3:2 --v4
with the subject going between 1988 and the dashes. So something like this:
red sky.png
would be DVD screengrab from a James Cameron fantasy sci-fi film 1988 red sky city coast --ar 3:2 --v 4

we're pondering the orb jerry.png
 
Last edited:
We did teach it Kiwis though, and I figured that would be cool to share since everyone here who's played with generating AI art probably knows how hard it is to get a decent kiwi bird.

View attachment 3960279View attachment 3960285

Is the model that makes kiwi available somewhere? would have been awesome to have to generate christmas cards a few days ago

these look great
 
  • Like
Reactions: SomeBirds
I hope when AI does take control of all creative endeavors it at least knows what POV means
Science is close; putting POV in your prompt can help with cowgirl gens.

Models and Stuff are compressed into a file format called ckpt. There can be malicious code inside which gets executed if you load it in SD or other programs. So always check your sources if the files are safe to download and use.
More specifically it's a model weights training checkpoint. "Model" is a slightly ambiguous term referring to part of a system or subsystem including the weights (it's not really ambiguous, it's just a more mathy definition than the colloquial one, and people often say model when they're only talking about the weights since you only need to exchange weights in a system to make it a new model).
Weights are basically what they sound like, just big arrays of decimal numbers applied to nodes in a (neural) network to influence the outputs. Outside of system design, literally all training a model involves is tweaking these weight values over iterations by propagating penalty values from some error (loss) function. You can pause and pick up training these weights again later which is why it's called a checkpoint.

Pickle is a python serialisation format. To get pickled is to get a malicious payload, it's a joke term applied to checkpoint files since they're similarly unsafe. I dunno if they literally use Pickle (I just use various ML library training wrappers to save weights, they might use Pickle underneath or not, I've never cared to check) but I don't think it matters.
There's various safety features in WebUI already and an unpicklable (and faster) format called SafeTensor is being introduced but I don't think it's all idiotproofed yet so yeah, be careful. Don't download and run random models anon uploads to Mega.
 
Back