AI Generated Images

[Anapan]

I've been seeing some pretty impressive AI created images lately, so I decided to try it out myself. I downloaded a GUI driven version of Stable Diffusion that runs entirely on my computer. It uses my GPU to run the heavy calculations. I spent a couple of days trying to make it create believable terminator cats, and I'm really impressed with what it spit out. Here's a galley of the highlights from the 700 images I had it create:

Robot Cats
I read the reason it's creating many pictures with two heads or too many/too long legs is because I'm having the AI render the images at 768 pixels instead of the 512 it was trained on. It isn't sure what to do with the extra space so it creates more things than the single object it would normally create in the space. I keep it at 768 because the loss of detail at 512 pixels is not worth the occasional conjoined-twin mutants. Each image only took about 1 minute to render.

[opa]

Terrifying and amazing
You must obey robot cat overlords

[Note]

Appreciate you sharing this, Anapan! I heard of AI generated images recently, but had no idea how they were being created, so thanks for sharing the info about Stable Diffusion. If anyone has experience with an app like this that would work on an Intel Mac, let me know. I'd have a good time messing around with this type of software.

The Terminator Cats are definitely a bit scary! Lol.

[marurun]

Why do so many of those cats have extra limbs? It's freakin me out, dude! It's freaky that some bits are so well-rendered in crazy detail and then there's just a random extra leg.

[Anapan]

@Note: There's some links under this youtube video for some browser-based implementations. https://www.youtube.com/watch?v=qATqEehWzzU

@Marurun: It is mostly because I've forced the AI to render the images at 768 pixels instead of the 512 resolution the AI was trained to create pictures at. There is so much more clarity and detail in the larger images, but the AI often can't reconcile what to do with the extra space and so it creates duplicates of parts of the image - faces, legs, random shape made out of the same texture as the limbs etc. I read that this issue is being actively pursued. This is a relatively new AI engine that is constantly being improved upon by many people since it is open-source. I cherry-picked the images for the gallery; All the hundreds of remaining renders are much less believable as cats or robots, and many are just a mess of slightly recognizable shiny shapes. I've considered photoshopping the renders, even extracting the good parts of abominations to create a good image, but I think the time is better spent trying to create more.

I'm not sure which of my search terms are fully understood by the AI. I looked at some examples of good renders and was just winging it - you can see my search terms and settings in the file names. The people who developed it have written a lot about what's actually going on behind the code, but trying to understand it is a little beyond my attention span and comprehension. I still think it's amazing that these start as colored static, and seem to be somehow built like the "IFS fractals" I used to play with in a DOS program.

If anyone uses an AI to create a cool thing, please share it. I'm always happy to host people's files on my site as well.

I found a search engine to find AI pictures made by other people:
https://lexica.art/

If anyone wants to see some more amazing AI and 3D-CG implementations, I've been following a YouTube channel called Two Minute Papers that visually explains some other amazing projects.