#let's not even mention Craiyon

[leam1983]

AI Art Shenanigans

I like to use Midjourney to visually workshop 'toons I can't get a lock onto, outside of their personality or motivations. I figured I'd share a few weird or interesting outcomes.

Long-Ass Post Below. Clicky.

And no, I'm not reopening the AI Art can of worms, using Midjourney doesn't make me a visual artist, it just makes me a dabbler in the greater field of pictorial art. It's about at the level of being three years old and making gouache splotches on canvas, except the splotches at least look like something.

Let's open with the best one I've got, which I've been calling The Diplomat.

This is the result of my realizing that Midjourney really likes repetition. That post I made about butchering a description of the Mona Lisa? I got to this level of detail by taking on fifteen superlatives to "older mustachioed man smoking a cigar". Some of these made sense, but others were oddly specific. How does adding fibre optic as a non-specified detail help with overall image composition? I don't know how, but it strangely does.

Here's the four source images that led to the first one on the top left earning the right to an upscale. It's there that I noticed that in AI terms, upscaling isn't so much a question of resolution as it's a question of detail. You'll notice that the Diplomat has two missing medals on his lapel, a more extravagant 'stache -

And horribly bungled fingers. I've tried Midjourney, DALL-E 2, Wombo and a few others, and all of them are uniquely bad at rendering human hands convincingly. In the best cases, the first five digits are realistically positioned, but the AI tacks on sixth or seventh ones because... I don't know. Humans = Fingers, maybe, so it emphasizes that somehow?

Asking for a full-body shot of the gentleman in question gets me what feels like a crossover between Ebenezer Scrooge, a Jojo's Bizarre Adventure Stand and an Attack on Titan baddie, as reviewed by Clive Barker:

That's - how many abs, Midjourney? And what's with the fingers, again? I count seven fingers on the exposed hand. SEVEN! Props on the design, otherwise, as it's a Very Anime Old Man, with a touch of Cenobite Kink, which is unique - but not entirely what I'd asked for. I don't hate it, but it's wholly unrelated to the source image, no matter how many re-renders I went through by pumping up the source's Image Weight value. For the uninitiated, "Image Weight" tells Midjourney how much an expected output needs to look like the input. The max value is 5, and I'd set it at 5.

It could be worse, though. MJ could more or less shit the bed and decide to go Diablo on the source:

Yikes. This guy has all the fingers and looks like he flays innocent souls in some -nth Circle of Hell for a living. He's probably uptight and sounds like Tim Curry at his pre-stroke gooiest.

Let's change topics. I've got Christmas on the brain, so let's spawn some Santas...

I particularly liked the ones on the right column, so I had these upscaled...

I'm amazed at how MJ takes to the basic concept of "Santa Claus" and manages to not only keep to the idea, but also riffs off of the shape at the first generation level! The one up top doesn't follow the expected color scheme, and you can still clearly identify him. The one below is a bit more traditional but, once again, Midjourney struggles with a few details. Santa's hands are a mess in both versions, and his toque's white pompom is just floating on the back of his head, in the second one. The first one keeps its lit.. spindly cigar thing somewhere below where his mouth should be, and the second one seems to be chomping on some sort of small wand while holding his stogie.

Insofar, I'm left with the impression that AI art isn't the big boogeyman we've heard about; not when all the pieces I've shown could benefit from human-piloted touch-up efforts - and an extensive re-training of its model on human extremities...

What's really interesting to me, though, is that The Diplomat is the product of my searching for an elegantly,dressed Mind Flayer smoking a cigar. I'll grant Midjourney's team the fact that Illithids are Niche Nerd Shit and that it's not totally absurd for an AI model to confuse facial tentacles for a mustache; but I would've liked something a little closer to the prompt. As said above, though, I still really like the final result.

And for comparison's sake, here's DALL-E 2 taking a crack at the same exact prompt. Note the more plasticky render - and its ability to render an actual squid-face, for once!

So, do I think the artists of Tumblr are in mortal peril? I'd say they aren't - most artists I know do a better job at drawing hands than any AI has, even when they falter or stumble in the process. For those of us who can't draw, though, this is a pretty useful tool to get ideas flowing.