I don’t dress my cat in costumes because, without even trying, I know she would hate that.
But now I can use text-to-image generators like DALL-E2 to imagine what she would look like in costumes. After all, even if it never saw my cat in a robot costume in its online training, it’s seen cats, it’s seen robots, and it’s seen costumes, so it can try to put them together.
Everyone knows that squares = robots so this costume makes perfect sense.
When I try to get DALL-E2 to dress my cat up as other animals, one common failure mode seems to be… this.
It has accurately put my cat into a plush giraffe onesie complete with hood, cuffs, and zipper, but it has also turned my cat into a giraffe.
There’s a similar failure mode with its tortoise costumes.
(It has also taken the libery of adding a pumpkin to the costume, I see.)
Here’s what it did for a scorpion costume.
It has picked up enough about “scorpion” to know that it should start doing insectlike stuff, but it’s throwing in ears, cow horns, bee stripes, wings, everything but the iconic scorpion tail and claws.
And here’s “caterpillar”.
Its antenna game is strong, I’ll give it that.
Another category of costume that gives it trouble: things with horns and snouts.
Its triceratops is spikes everywhere, plus a curly unicorn horn.
Its elephant is also somehow a unicorn.
Its anteater, which shouldn’t even HAVE horns, is also a unicorn. And the cat is some kind of weasel?
This is also an anteater costume. Here an anteater appears to also be some sort of quail in a sweater.
DALL-E2’s tapir costumes were also really weird.
Clearly a tapir.
It’s not that it doesn’t know what a tapir looks like. Here’s its attempt at a tapir in a tapir costume.
I’ve previously noticed that something is weird about how DALL-E2 deals with multiple kinds of animals in an image, usually choosing just one kind of animal to generate, or blending the two. It seems a similar kind of weirdness might apply to animals dressed as one another.