A lot of hype has raced about the internet recently concerning AI art and how computers will replace artists. Fearful painters and illustrators need to start searching for new jobs because creativity is now only for algorithms!
Silly, isn’t it?
I’m a fantasy/dark fantasy artist. I began with good ol’ pencils, pens and paper, and transitioned to a pen display because I wanted to save trees (and erasing is SO much easier and I love layers). I’m not the only artist that made the transition, and I know that choosing to create electronically did not eliminate watercolorists and oil painters and sculptors.
Erse Parr, artwork I created for my Evenacht serial. It’ll be a while before AI “learns” enough to output images that aren’t front-views of staring individuals in front of painterly landscapes.
So it is with AI. AI lets non-artists create pictures they can use in their creative endeavors. Great! Especially for independent writers, DMs, etc., paying for a pic is often out of reach due to budget constraints, and websites are popping up all over the place that will let them make a landscape for free.
But that does not mean artists are going extinct. And they never will. There are several reasons for this.
*Before I get started: I use Artbreeder regularly, have messed around with Night Café and NVidia Canvas. I’ve watched how-to videos on Midjourney, but have not personally used the Discord app because I initially thought the damn thing was a scam to swindle personal information from users (have you looked at their website? But hey, you can pay $30 for the privilege of beta testing for them! Ain’t that grand). I’ve fiddled with other AI art apps, but being unimpressed with the output, I stopped using them–and I really don’t remember much about them.
1. AI cannot give you exactly what you want. With computer-generated images, you need to nudge sliders, add keywords, backtrack, and even then, the image achieved may be nowhere near what you originally intended. What if you love a part of the pic? Changing anything will change that, too.
Artists know what they want, and can create the image they picture in their heads. If they like something about what they’ve drawn or painted, but want to change another aspect, they can do so without creating an entirely new image (MidJourney, Night Café) or warping what they like because something else is off (Artbreeder, NVidia Canvas).
Pumpkins created in Night Café. Key phrase: The left is “pumpkin with glowing eyes and wings sitting on a box” with modifier “horror Gustav Doré Greg Rutkowski”. The right is “pumpkin with glowing eyes and wings sitting on a box” with modifier “sinister by Greg Rutkowski”. As you can see, some of it is cool–but not what I wanted. How many keywords do you think it would take, to get a pumpkin with glowing eyes and wings sitting by itself on a box?
If you hire an artist, you can converse with them about what you like/don’t like about an image and get something close to what you originally pictured. You don’t have to spend hours attempting to get black hair back on your character without having an eyeball appear in their cheek or a hole in their neck.
2. But! wait until AIs learn how to properly create a human face! Then artists will be in trouble! Of course, algorithms are notoriously by the books. Once a program “figures out” what a human face looks like, that’s what you’ll get, with little diversity other than eye and hair color, let alone style.
Amderion, a Finder sage from my Evenacht serial, created using Artbreeder then transformed in Photoshop. Trying to get long hair on him without destroying the rest of the look was hours of frustration. I should have just drawn him because it would not have taken as long. I also find it interesting, that a couple of commenters did not think the portrait looked very Artbreeder-y. They’re right, I suppose; Amderion isn’t a younger woman with chubby cheeks, full lips and wide eyes.
Let’s look at Artbreeder portraits. I use the ones I create as placeholders until I find the time to draw the character (sometimes that happens, sometimes not). Their images are immediately recognizable to my husband (who is not an artist and has never used the program) because the results are square portraits that display the head, possess a dull background, and have pretty much the same appearance. Smiles and frowns tend to warp the face, so nearly all have an intense “stare” look ( and have you tried to get face-front portraits to look up, or down, or 3/4 or 1/2 view? Or *gasp* look like a person over 60? Not pretty results).
Yes, Artbreeder has a re-roll feature, which is interesting but after you’ve produced ten or so re-iterations, the re-rolls tend to look like the previous ones. And what about long hair on men, or short hair on women? As soon as you push the slider towards a gender, their hair either lengthens or shortens. Since my characters tend to have long hair (all of them. I like long hair, what can I say) this drives me nuts, which is why they get plunked into Photoshop for editing.
This is a huge problem, though. It means the algorithm is learning a strict set of gender-related cues, which, again, limits diversity in characterization (and which produces a very warped sense of ‘beauty/handsomeness’).
3. Artbreeder isn’t the only program that works from strict sets of cues. All AI does. Due to copyright restrictions, the creators of the algorithms are restricted in the inputs they use, which limits the output (or they break copyright and hope no one can prove it). Using these inputs, once the algorithms decide what a desert or a column is, that’s what they will produce.
Let’s take deserts. The Mojave Desert is very different from the Sahara, but I have yet to use an app that creates Joshua trees rather than dunes to depict a dry landscape. Earth has a wide range of biomes and an astounding diversity within them, and AI misses that.
NVidia Canvas isn’t too bad when it comes to landscapes, but that’s likely because you can paint where you want things to go. It still requires a lot of fine-tuning and lacks detail. I originally wanted a tree-lined escarpment overlooking the sea. Fine-tuning produced this. I like it, but it’s not what I had in mind.
And, of course, even if an artist does not hold copyright on an image, they can use it as a visual reference as long as they don’t copy it. Apps like Midjourney legally can’t, though I’ve watched tutorials where the user finds a Google image and plunks it into the data set, whether legal or not. This could be a huge problem in the future if the owner of a pic decides to sue for infringement because the result looks too much like their copyrighted image.
If you think that’s unlikely, you might want to read about Shepard Fairey, who was sued by the AP for using their photo of Barack Obama in the famous HOPE poster, and who ended up fined and put on probation for hiding evidence about it. Or how about reading Cory Doctorow’s account of harassment by underhanded lawyers concerning Copyright Commons (basically, businesses search for breaches of copyright and demand compensation from users of images etc.–and get this, they don’t even have to represent the original copyright holder. They can just accuse and reap the monetary benefits because those they target usually can’t afford a lawyer).
Or you can look into how YouTube creators get copyright claims against them for doing typical filter sweeps in synth showcases. Or how about these two louts who stole $23M from leading Latin artists because YouTube’s copyright enforcement is shit. The lessons here are: anyone can sue you for unauthorized use, even if they are in no way related to the original artist, or if the original artist is in no way related to the use. And yes, that sucks.
4. Which leads us to another copyright problem–the artwork itself.
In the US, our laws have not caught up with technology. Images created by these programs are not copyrightable, and truthfully, considering how many of the images look like thousands of others, who wants that? Having users suing each other because the portrait of their book’s main character is exactly the same as the main character of their rival’s book, only with blond hair instead of brown, is not exactly what the coders of these algorithms want. Such suits will kill their app in a matter of months.
Artbreeder background, human figure created by me in Photoshop. This background is based on a common landscape base found in Artbreeder, with thousands of iterations used for a variety of purposes. With so many variations on the same image, does any one of them count as the “real” one?
5. So you still think AI can replace human creativity?
The reason these data sets produce weird, interesting results is because the algorithm hasn’t figured out how things “look” yet. All that creativity will die once they do because, as stated previously, computers are literal. So while a lot of people like the monstrous looks Midjourney or Night Café produce, those will go away once deep learning “learns” and uniformity takes hold (which I see in Artbreeder).
AI is a tool, no more, no less. Computer-generated images can be an interesting first step in the creative process, but as of now, the end results will not replace artists–and I see no future in which they will.