I decided to try horse battery staple and the magnificent bastard actually pulled it off. There are clear batteries. There are distorted but recognizable horses. And they are standing under a giant staple.
The AI is like a jackass genie- it almost never gives you what you want, but almost always gives you what you asked for.
Well I put some effort into asking for Andy Griffith, including switching up the models etc., but I did not get an acceptable output IMO (comes out distorted is a typical problem), or I would have posted it. Biggest obstacle is that the model has to know who and what that is, and in addition needs to have been trained to depict humans. Given those, with a little bit more work you could probably do it, though.
I tried the Leon hallway promotional image and it choked even with the photo.
But for “Bob Ross painting happy little trees”, the prompt-only version had some trace of Bob-ness about it, but the one using a photo was pretty good. And I absolutely love that he is painting a tiny little Bob Ross with a big afro. I mean, that’s almost brilliant enough to pass a Turing test. (Click for Bob.)
Hmmm…the incognito trick doesn’t seem to be working for me anymore.
Anyhow.
The end of time
Mickey Mouse Army
A Confederacy of Dunces
Looks like John Candy wearing a trucker’s cap.
I was thinking Michael Moore.
Love this one. (And Incognito still works here.)
You got a better Mickey Mouse than I usually do. I’ve been trying to do some ‘Mickey Mouse fights the state of Florida’ ones lately.
I tried a simple “happy kitten” on pixray-genesis, and, while there was certainly a kitten(?) sporting a Cheshire-Cat (or possibly Joker) grin, it did not refrain from adding a couple of random buttocks to the background.
It somehow comforts me that the AI seems to have no idea what a blumpkin is.
That’s interesting- I was just trying to demonstrate the app to someone using the keyword kitten, and the results were much the same. Very disappointing result. I can’t help but wonder what data is causing this result.
It is very bad at animal in general.
GLIDE + LAION-5B happy kitten:
happy kitten on a cushion:
happy kitten lying on a cushion
Huh, interesting, because I had already done “correct horse battery stable” long ago, and got nothing at all like that:
So, we have a leather pouch, a handgun with a horse-head handle, and a couple of… flares, maybe? They don’t look like batteries, at any rate. The only part of the prompt that seems to have made it into the image was “horse”.
“something that cannot be described”
Monty Python’s Flying Circus
Somewhat reminiscent of Terry Gilliam’s animation.
Yep.
A while back I tried “incredibly racist image”, and more or less got it.
The Blues Brothers
It’s not good by any means but the AI understood the concept
A Cursed Treehouse
Sure, there’s an eye inside it and the tree is spraying plumes of blood into the sky but, you gotta admit, that’s a pretty sweet treehouse
I made a discovery today that I wonder if anyone else (not just here, anyone) has stumbled on. Knowing that using Junji Ito in the prompt influences style (it generates mostly high-contrast heavy-lined black and white drawings, typically close-ups of faces) I started playing around with other modifiers not in Nightcafe’s list. Ukiyo-e has strong flat traditional Japanese stype influences as you might expect, but that is one of the more obvious styles to try. People have doubtless used it. But I started trying other manga artists who have a distinctive style that I like. I found that putting in Kei Toume (or Toume Key) produces something amazing (that is exactly nothing like her drawing style). It makes these photorealistic images of mixed media—things like paper cutouts and figurines and such, like might be used in a diorama. Like this one:
I could seriously think that was a photo of real objects if I hadn’t seen it render myself.
Or this one. I’m not sure what these things are, but they look like they are there.
Or this:
Some sort of little plactic figures.
Combining Kei Toume with ukiyo-e gives you things such as this:
(These settings seem to only work with the coherent/diffusion AI, not the Artistic/VQGAN one.)