Some Thoughts on AI Art
Posted on August 11, 2022 Posted by John Scalzi 16 Comments
First, this one is called “Deep in your heart there is a sunlight so hot that it makes you love people. That’s why you love people,” and its inspiration is a little poem my daughter wrote on her sixth birthday. I gave the whole poem to the AI art generator Midjourney as a prompt, and this is one of the things it came up with. It’s certainly evocative.
Since Midjourney, Dall-E and other AI art generators have come online, there’s been a bit of a freakout from actual artists/illustrators about what this means for their livelihoods. While my own prognostication skills are dubious at best, and I would never tell anyone not to be concerned about the creative sector they work in when new technology surfaces, in my experience of using several of these AI art generators over the last few weeks, I’m not sure I see them replacing human illustrators to any great extent any time soon. This is for several reasons:
1. Specificity and intentionality: One can prompt an AI art generator in the direction one wants them to go, but ultimately you get what you get with them, unless you really want to devote a lot of time to art directing the thing. It’s still easier to communicate what you want to an actual human and get an exact result, than to go through 25 iterations of an idea and hope the AI finally gets what you want, without messing up anything else.
2. Detail: Most of the images I get out of AI art generators are of a level that I would call “cool rough draft,” which is to say, there’s enough there that you see where it’s going, but the detail level isn’t there, and what detail is there is wonky. This is most notable with human facial features, and shapes of distinct animals and other natural objects. If I were wanting to make the image above into an actual piece of art, I’d hand it over to an artist to get it to a level I would considered finished. I think at this point AI art generation is a handy way to sketch ideas and concepts, and for someone like me would make it easier to let an actual artist know some of what I was thinking. But the handoff to an actual artist would still need to take place.
3. Sameness: Having played with several AI generators now, I can say it seems each has what I would call a “house style.” Midjourney, which is the one I’ve played with the most, has a distinctly “arty” and “moody” style that I think I would call Emo DeviantArt. I like it! But I also know, barring very specific instruction, what I’m going to get out of Midjourney when I give it a prompt. Which means even two weeks in I’m getting the feeling I know its default bag of tricks. Humans also have their own styles, to be sure, but also more flexibility. Human work feels, how to put it, less programmatic.
AI will get better at generating art — the amount to which it is better now, at effectively its second generation, from its first generation, is a really actually impressive — but I suspect it’s going to keep bumping up on these problems, because “AI” isn’t actually intelligent in way a human is, which will continue to give humans an advantage on generating art other humans actually want.
What I suspect is going to happen is that human artists will start incorporating AI art generation into their tool box, and that very rapidly; if AI can, for example, quickly generate a background cloudscape that is consistent with that artist’s style and intent, which that artist can then tweak to suit their needs, why wouldn’t they do that? Saves time and the final work is still under the direction of a human brain. Likewise, in the next generation of artists will be some who can’t draw to save their lives but who are maestros of prompting art generators to give them things that no one else can get out of those generators.
And for people like me, who have very little visual art talent, these AI art generators will let us play a bit and perhaps will spur creativity in other directions. I’ve already created some images that I want to write stories for, or which have at least have ideas popping into my head. Will anything come of those? Maybe, maybe not, but it’s nice to feel the creative ferment they help create.
So, no, I don’t suspect AI art generation is the end of human artistry. It’s another tool we can use, and I think it will be interesting to see what happens with it as we go along.
An artist I know has already been doing some cool works incorporating these, like a bunch of alt-70s prog rock album covers, and a video made up of edited still frames (possibly from more than one AI). I’m sure they’ll quickly become another tool that most digital artists use for something — I think it’s going to be a very long time if ever before they significantly ‘take over’ the jobs of any human artists.
If the good AI art generators allow you to use the work freely, I would tend to see this replacing a lot of the “we just need something visual to go with this [blog post/newsletter/etc.]” stuff in cases where people actually care about copyright/legality. So: not going to replace beautifully hand-crafted illustrations, but reasonably likely to replace a slightly-paid sector of stock photo/illustration, provided the AI is cumulatively cheaper.
I guess, I could easily see, say, real estate agents who blog and want art of a welcome mat (or of keys, or of a fireplace, or of cookies) turning to this, if it’s cheaper than a stock photo/clip art subscription. Exactly how much of the market that would be, I don’t know; and I also don’t know how high into the market for illustrations it’ll go, but if it’s copyrightable (such that the user is getting something that they can use and their competitors can’t copy off their social media posts or whatever), I would expect it to shake up the “we just need something to go with this text” sector of the market. But a very large part of that sector of the market is not giving much to artists anyway – they’re paying peanuts to clip art and stock photo companies (or just grabbing stuff from the internet without worrying about copyright, depending).
I’d rather art and graphic design not get more precarious. But I’d be surprised if this were not adopted by places that are already heavily cutting corners with graphic design. (generic-brand tissue box art! small to medium businesses and nonprofits which are not art-centric but which need some sort of images for their social media and advertising! Walmart shower curtains! Low-cost patterned fabric!)
I like your digital art better than I like Picasso’s work, so there’s that.
It’s more recent, at least.
I’ve been playing with Midjourney and have come to similar conclusions. If you want something goofy, abstract, or surreal and aren’t looking for a specific composition, AI-generated art is a quick and easy way to get something lovely. I’ve gotten several fantasy landscapes that I absolutely love, for instance.
If you’re looking for a specific composition or realistic details, and you don’t want to spend a ton of time tweaking prompts and/or photoshopping out the distorted details, then stock images or custom work from a human artist are going to be the way to go.
Many years ago I asked an artist friend to draw me a unicorn playing guitar onstage. I can no longer find this treasure anywhere in the house, and tried all of the AI art generators I could with that prompt and the results were…. not great. The best one was terrible and the rest were worse.
I’m a (retired) professional illustrator and I agree with Anthea, above. Some beautiful images, most with a fantasy feel (which I think influences my personal idea of beauty). Some you might even frame and hang on a wall.
But, as John says, there’s a lack of specificity needed for assignment work.
I am strongly reminded of 1994, when I got my hands on software called Bryce from the maker of Kai’s Power Tools. As Wikipedia says, it did 3D fractal landscapes.
I managed to make a SF con program book cover using it (Minicon 30). But my method was to generate a bunch of semi-guided landscapes and pick one where I could Photoshop minimal elements into it.
Nothing I could’ve used for my commercial clients. But fun!
More on 1994 for any who care. The first version of Photoshop to have layers was 3.0, released for Mac in Sept 94. And Minicon 30 was in April 95.
Did I use the version with layers? I was on a limited budget. I can remember doing some compositing of images without having layers functionality (because it was very, very painful), but I can’t recall this one specifically.
BTW, even more OT: If anyone here knows how to recover files saved in the 90s in Photoshop and Freehand formats, please speak up. Most appear as Unix executables now. (I later started saving every “final” file in a generic format like EPS and TIFF as well…. slow learner.)
Are these owned by you because you put the text into the AI, or are they owned by the owner of the AI, or are they owned by Athena and Robbie because they wrote the text you used?
Regardless, interesting and quite nice.
Q: was the generated unicorn image disappointing or horror or ugly? if horror then please post… just what the world needs now, a brutalized mashup of unicorn and music… just needs a forest burning onto ashes in the backdrop and it will be ideal as cover art for everyone’s 2022 high school yearbook…
I’m guessing this will further push the evolving art ‘n craft of e-books, since there’s no complex 4-color offset printing to deter beancounters at publishing houses and no matter how bloated the resulting file (3Mb? 8Mb?) no e-book could ever take more than 3 minutes to download… be cool to have customized imagery inside a fantasy novel… maybe a war weary unicorn…? a kraken with an eyepatch…? a vegan warlock sacrificing tofu on an altar…?
Bruce Sterling has been doing interesting things with Dall-E by starting with the prompt “crungus.” As an English-looking word with no previous meaning it should have produced fairly random images. But no, the crungus is something very specific, at least according to Dall-E.
That is something you should be able to control in the prompt. At least to some extent. I’ve seen plenty of examples of people using the same prompt but with qualifiers for one or more stylistic aspects.
“or are they owned by the owner of the AI”
OpenAI at least, has specifically said that anything generated by their tools belongs to the person who generated it. Which is nice of them, but doesn’t address the millions (billions?) of examples of other people’s work that was used to train DALL-E. I’m not convinced that those people don’t deserve some say in the matter, although that seems like a lost cause at this point.
Ursula Vernon has been doing some very interesting stuff with Midjourney: mostly backgrounds and other things to make it easier to play with ideas. She’s been posting some results on her Twitter; this is a sample:
Also a dreamlike comic #Goldfishcomic and an insect tarot: #TarotInsecta
“AI” in the middle of ordinary text without context looks a lot like Al, the name (as in short for Albert etc.), so for a moment I thought we were going to be reading about art done by people named Al.
@Howard, Ugly and poorly done (like misaligned parts of the body, clipping, etc). They largely had the aesthetic of cartoon furry porn (which probably says more about what the AI are most often asked to create than anything else).