Oh boy. I think that clears that one up. Thanks.
I tried another Modigliani Paris Hilton with some extra tags and qualifiers like “magenta iridescence” and it came out much stranger:
1 Like
Gremlins by Rene Magritte
6 Likes
Holy shit.
Run DMC on a waterbed
Looks like some Jewish painting my grandparents would have had.
5 Likes
AOC as pixel art:
1 Like
I made “Paris Hilton riding a unicorn, photorealistic painting trending on Artstation” and I think this qualifies:
1 Like
I made a discovery though. We are using the wrong algorithm with that VQGAN thing I posted. The better one to use is CLIP guided diffusion models . The 512 takes a while and probably uses more GPU than free, so here is the 256x256:
You can AI upscale it using this:
However, these two links here may be easier.
Try turning up the clip guidance scale. Like maybe to 1500 or 2000. Your text prompt matters a lot though. That’s the whole game here basically.
I’m using one of the very user friendly ones. Just typing in some words and clicking start.
Here’s “Muppet treble clef water color”
6 Likes
Isn’t this supposed to be a politics forum? You should use these sorts of new technologies to find answers!
‘The reason why trickle down does not work’
Is that Trump? 
2 Likes
…seed text?
“An oil painting of Jesus during Gay Pride week.”
Seems about right. They would absolutely string this dude up.
6 Likes
This is from Part 4 of a series on text to image. Some crazy examples here like that Buzz Lightyear pencil sketch.
Figured I had to do it.
“AI Art Machine”
5 Likes
By the way, I snuck the link in there but the AI upscaling thing is real and has been around for a while. We laughed at CSI for the “enhance” meme but now it’s a real thing, and it was a fairly obvious thing that was going to exist before it existed if you understand how ML / NN works. There’s plenty of free stuff out there that you can clone from github and try on your own if you have enough cuda cores, but one of the best known commercial examples I’m aware of (especially for video) is from Topaz Labs. Example here:
Now there’s a lot going on here with deinterlacing, denoising, frame interpolation to bump it to 60 FPS, etc., but this guy is definitely using Topaz to run it up to 4k (read the process in the video description). One aspect that really sticks out to me in the upscales is the treatment of specular highlights like the shimmer from that ridiculous gold foil jacket Jimmy Vaughan is wearing and the color shifting of the iridescent stickers on Stevie’s guitar.
Of course, none of these are going to be perfect. They are imputing values that don’t exist based on training data where values do exist, so it seems like the performance depends a lot on the training set and specific implementation. For instance, there’s a free still image upscaler called waifu2x for anime / cartoon style images, but it’s performance on photorealistic images is a bit suspect.
Anyway, that’s a long-winded way of saying “these technologies exist” for anyone who wasn’t aware of them. Generating high-resolution AI images is GPU intensive so people are doing the 256x256 and upscaling.
Here’s another one that’s easy to use. Text seed was “landscape.” Your kid could seriously not paint this:
3 Likes
“Build Back Better acrylic”
I’d hang this on my wall.
3 Likes
“Establishment Democrats” from Huggingface CLIP (10,1014,152,0,0,0,50,32):
2 Likes