AI Art Doesn't Replace Real Artists...Yet

Last week, Michael Offutt posted a link to this site called "Stable Diffusion" that lets you put in keywords and an AI program generates 2-4 pictures based on what you write.  I decided to try it to look for some things I was looking for on stock photo sites and then just some other random stuff.

The results were mixed.  Sometimes it comes up with something pretty brilliant and other times it comes up with something pretty horrible.  And sometimes it doesn't even show anything because it gets censored for being pornographic or racist or something.  Like one time I copied a description from a stock site like "woman in Santa hat with Maine Coon cat" and it got censored.  I think the "coon" part was prompting it to do something racist.  Weird.  Here's an early sampler of three searches:

And when I asked it to draw a warthog, it came up with nightmare beasts instead.  That seems pretty weird that it doesn't know what animals look like.  When I just put in "Pumbaa" it came out with weird creatures that were more cartoony but still creepy.

I did another little test, trying to describe the cover for A Hero's Journey to see what it would draw.  The four results and the original picture are here:

A couple aren't terrible, especially the first one, though they aren't really what I asked for.  So Rusty Webb 1, AI 0.

I think the best results come if you're not asking it to do people or animals.  It can do almost photo-quality landscapes:

For people I think it did better telling it to do a close-up:

Telling it to do a portrait would sometimes generate something like a painting, but it also came up with some decent results:

Sometimes it would do a mostly-good image, but it would do something weird with the mouth or especially with the eyes.  It's not always just "Uncanny Valley" so much as it gets kind of lazy and just doesn't do both of them consistently.

The larger one on the right looks like an android.  Of the other four a couple are pretty good and a couple, not so much.

No idea why it did the first one in mostly sepia but it's neat.  The second was good except for the "freckles" that are like a rash on one cheek.

I like the first one except for the purple and blue "freckles."  The second the eyes are a little weird.

The interesting thing to me is that it was kind of like searching stock photo sites.  So many times to try to find what I want, I have to plug in keyword after keyword to find something I want.  In the same way you might have to keep changing your keywords to get it to generate something that's usable.  Like with stock photo sites, it can be really hard for it to get what it is I'm looking for.  So I often have to keep trying different phrases, hoping that one of them will unlock a decent result.

Like I started looking for an "African American female superhero" and it gave me weird cartoony drawings--a couple maybe a little racist.  Then I asked for something like "Black woman in mask" and it drew black women in Covid masks.  Then I changed it to "Black woman in superhero mask" and it came up with a couple of decent ones.  So like with stock photo sites, you have to refine your search terms until you maybe hit paydirt.

Obviously it's not going to replace artists right away, but it's probably a start.  Before it can really replace humans it needs to get a little smarter.  And if you wanted it to do comics or anything like that it'd have to be able to make things consistently; you couldn't have the thing spitting out completely different looks of the main character for every frame.

Anyway, since this is free to use for right now it does add another possibility for when I'm looking for something to use on covers and such.