19 comments

[ 2.6 ms ] story [ 53.7 ms ] thread
This is great! It's fun trying to decipher the various furballs of different shapes.

(Some are reminding me of _The Thing_ but I'd also probably pet them)

"Enterprise, what we got back didn't live long"
Maybe they do exist. Unless one has encountered all cats past and present, it is impossible to know of the existence or non-existence of such cats as situated in this web place.
Fully half of mine are nightmares. Disassembled cats. Multi-headed floof piles. Headless flood piles. A pile of kittens stitched together. Heads on the wrong end. Cat with kittens emerging from it's marsupial pouch.

And some are cute.

All are far better images of cats than I could draw myself.

Sometimes these look like cats on dog bodies
(comment deleted)
Every single one I get looks odd. Even the coherent ones. There is something about the eyes in relation to the rest of the face somehow related to pseudo depth perception.

Can you put it into words?

this gets the feet almost right!
I saw one very normal cat picture come up twice. Has anyone done an an analysis on the underlying data to see how unique these “do not exist” pictures really are? Sometimes I wonder if these are little more than stiched together collages, and some times made up of very few source images.
https://d2ph5fj80uercy.cloudfront.net/06/cat636.jpg https://d2ph5fj80uercy.cloudfront.net/05/cat2788.jpg

As far as I can tell, identical fur patterns on the face, but different aged cats. Interesting.

Yeah, maybe I’m the outlier but I don’t know if a lot of the AI stuff that’s coming out right now is doing what people think it’s doing. My impression is that the AI is learning the minute details that make up a cat and confabulating something completely new based on those details, not stitching together images of existing cats. If your three your old cut out a picture of one of your dogs on a picture of another of your dogs I don’t think you’d think that was analogous to what this AI is doing, but I think it’s basically that (maybe we a dozen or so pictures but seemingly much less than my gut would indicate). But more of a collage than an artist using a combination of their memory and imagination to draw something completely new.
When I get into obscure areas, I definitely get the sense we're remixing a small library of basic components (I've been generating variations on "19th century hand painted wallpaper" with stable diffusion recently). In this niche I have less and less confidence I'm not just getting someone else's work that's been put through the intellectual equivalent of a bitcoin tumbler.

But I don't really know enough about the this particular black box to have an informed opinion.

Maybe I'd get broader results with one of the commercial players, but maybe that's only because they have a bigger library of basic components.

Actually, you know what? We know it's just a remixer, because sometimes the shutterstock logo peeks through.

I wish you could see a sort of RGB gradient output of the stitching that occurs and observe the patterns between good outputs and dreamlike nightmares.

Something like a normal map or UV texture, you know? But instead it’s just an RGB output with solid colors representing individual source blending or contextual blending of some sort.

Thankfully these cats do not exist. I'm just seeing nightmare floofs.
Sad there were no torties. Maybe it's never met one!