Monday, June 20, 2022

Art For Art's Fake

Late last night, Wilhelm alerted me to an internet fad that's apparently been all over Twitter for a couple of weeks. Not that I'd know. I'm not one of the Twitterati. 

Maybe I should have known, even so. When I followed the links, first to the version of Dall-E that Wilhelm favors and thence to Craiyonits new home, I realised the whole thing was part of something I already have bookmarked, the GPT-3 Demo Showcase.

As with a lot of aggregators, there's just too much to wade through there. Over three hundred AI-related apps using the GPT-3 large language module made publicly available by Open AI last year. I'd been riffling through a few of them in idle moments, playing around with various "writing assistants" and chatbots, but I'd been under the impression everything there was purely textual.

It seems not. Craiyon is an "AI model that can draw images from any text prompt", something like the Artflow app we all had so much fun with a few months back. Except that where Artflow only generates portraits, Craiyon attempts to visualize anything you suggest. 

It's not content with a single try, either. Craiyon gives you nine different versions in a classic comic-book grid, something which creates, in my mind at least, an almost subliminal expectation of some kind of narrative progression. It adds to the general surrealism of the process, as if it wasn't surreal enough already.

Even though it was late, I couldn't resist playing about with the thing right way. I did a couple of runs at the Dall-E Mini site before moving to Craiyon, beginning with this one:

I wasn't entirely convinced. Most of the shots look a lot like Tony the Tiger but not much like they've been drawn by Bill Watterson, the creator of Calvin and Hobbes. #7 looks something like Hobbes. I'm not sure any of the others do.

I was curious as to why several of the shots have distorted, almost cubist faces. Nothing in the seed should have suggested it. As things progressed it became obvious at least a degree of facial distortion was just about guaranteed, no matter what seed was used. I still don't know why. Maybe the AI's on something.

Still on the tiger trail, I tried Shere Khan next. Once again, I couldn't really see much - or any - of Bill Watterson's influence in the result. I started to wonder whether Craiyon really knew who Bill Watterson was or what his drawings looked like. Of that, more later.

There followed a very unsuccessful series of attempts featuring Sabrina the Teenage Witch and Gidget. In only one was either of them even vaguely recognizeable. This one.

The AI seems to think "various" means "identical" but it's very keen to prove it knows what a witch is. In # 4 it seems to have given Sabrina two hats to wear and by the last frame it has at least managed to come up with a color other than black. Still, not very impressive.
Back to the drawing board, by which I mean more cartoon characters. I was wearing my Stience T-shirt so Bean from Disenchantment was the first one I thought of.

It seems that being the lead character in a successful Matt Groening animated series is no more a passport to fame than landing a long-running role as the face of a once-feted mmorpg. The AI had no more idea who Bean was than it recognized Firiona Vie's name when Wilhelm pointed it in that direction. Of course, the fact that I got the name of the show wrong (As I nearly always do.) might have had something to do with it. 
[Edit: I just reseeded it with the correct name and got this:

My apologies to the AI. I am an idiot.]

It clearly knew what a  kite was, anyway, so I tried again with someone I thought was bound to be famous enough to be pictured flying one: Elvis

I really like those. It's nice to see Elvis looking so happy. 
Since the AI was clearly at home with Elvis, I thought I'd give it a chance to show off. It was plain from Wilhelm's post that one of the app's great strengths was cloning the signature styles of famous artists. If Bill Watterson wasn't famous enough, how about Banksy? The internet loves Banksy.

Wow! That's just about perfect. It would make a great poster. Or a mural.

It was at this point that I had a brainwave. Maybe Craiyon didn't know who Bill Watterson was but how about his most famous creations? How would it be if I asked it to draw Elvis not in the style of Bill Watterson but in the style of Calvin and Hobbes?

Oh boy. It loves Calvin and Hobbes! So much so that it barely remembered to put Elvis in at all. That's not "Elvis in the style of Calvin and Hobbes" so much as Calvin and Hobbes with a very brief cameo by the King.

Since Craiyon was so into them, I gave it another chance to indulge itself.

Apart from the distortion, those are very good fakes. It wouldn't be all that hard to add some speech bubbles and turn them into a strip. It's easy to see the potential for mischief, something Calvin might love but I'm pretty sure Bill Watterson wouldn't.
I thought maybe it was time to move away from cartoon characters and copyrighted material. Since Elvis worked so well, how about a couple more rock stars?

I didn't specify "...during a pandemic" but Craiyon decided to add masks anyway. Other than that it's a mind-blowingly accurate visual record of an incident that never happened and probably never could have happened. I'm particularly impressed by the way it correctly interprets the preposition in all but two of the pictures.

Time to step down hard on the surreality pedal. I'd been keeping my texts short and to the point. I decided to go long and complicated instead.

Once again, I was impressed. "Ann of Cleves" doesn't feature but then I did spell her name wrong. I guess you can't tell it's a Thursday, either, or what they're talking about, but it's definitely night-time, they're obviously having a discussion and that could as well be The Factory as anyplace else. The rabbits, the monkey, Nico and Andy are all clearly recognisable, even if facial distortion is at an all-time high.

It freaked me out a bit. It's like set of contact negatives from a photo-shoot for Interview magazine, one that was judged too disturbing to publish. It was getting properly late by then so I decided to call it a night before things got any weirder.

This afternoon, just before I started on this post, I ran a couple more ideas past Craiyon. Rather than leave you with the unsettling images above, here's something decidedly more cheery. And colorful. Boy, is it colorful!

Doesn't that have a weird Rembrandt/Vermeer feel to it? Ferris looks as if he's wearing a ruff in #2 and a cassock in #7. There's an awful lot of Scooby and not so much of the gang. I can make out Daphne in #3 and Velma in #5, I think, but if Fred or Shaggy are in there anywhere, I'm afraid I can't see them.

Saved the best for last there. I'd have that on my wall, blown up to poster size. That and the Banksy. I love the way Scooby seems to be wearing clothes in most of his shots.

As I said about Artflow, this would really integrate wonderfully with Occupy White Walls. I really ought to look into that. I think it might be doable. Whether or not it would be legal is another matter altogether.


  1. Now I am going to have to go re-run a bunch of my favorites through Craiyon.

    Both do seem to have a problem with faces, even when you use light art styles. I had "courtroom sketch of Richard Nixon cross-examining a jar of pickled eggs" and, while I got a sense it was Nixon, it was still a bit off.

    1. I wasn't 100% sure whether Craiyon was a different version of the AI or just a rebadged copy. The stated reason for the change appears to be a request from Open AI to rename it to avoid confusion with the original but it wasn't clear to me whether there had also been a substantive change to the software.

      Either way, it's a fun toy but having seen what the big version can do I'm a lot more excited for the eventual public release of that.


Wider Two Column Modification courtesy of The Blogger Guide