The title says it all. There’s an abundance of free AI art programs out there, capable (or not, as we’ll see) of producing amazing artwork. The three main factors that determine the output seem to be:
- the pool of resources the program has learned from
- the quality of the program
- the text prompts
Now, when this stuff started to come out, people started to panic that computers would take over creative product, followed by a brace of cynics who insisted computers could never replicate that extra-special singularity possessed solely by humans, Real Creativity. Then AI art programs improved, competitors and imitators emerged, and it seemed that, yeah, computers could be wildly creative with their product.
The next issue to come up was artistic integrity. Maybe computers wouldn’t take over all creative product, but in order to inform their product they were being fed databases of works by real-world artists, both dead and currently alive. That enabled me to make artwork by artists who never produced this kind of art, namely, a Black office worker done in the style of Edward Hopper.
This was an interesting exercise! You could take any subject you like and render it in the style of an artist who was no longer creating work. Make a portrait of yourself in the style of Johannes Vermeer, render your yard in the hand of Camille Pissarro, see your pets through the eyes of Jan Brueghel the Younger. Why not?
The logical issue here was, if a deceased artist, why not a living one? Now we’re encroaching upon people’s livelihood. It’s one thing to ape the distinguishing marks of a famous former artist, but to create new work in the style of someone who’s still alive and trying to earn a living by their craft, that seemed like a violation. When you enter your prompts for subject matter, to create an image, you can enter other influencing factors like high resolution, dramatic lighting, complementary colors, or having it done in the style of Greg Rutkowski, a Polish artist specializing in epic and dark fantasy. Many AI programs not only offer his name as a prompt, some have him listed as a button to click. “The highest form of flattery,” indeed.
Note that the article I linked to was written by Melissa Heikkilä, who simply wanted an AI program to enhance a selfie and received a barrage of lewd, lascivious material as a result. That’s another issue: what are we feeding our AI engines? I can’t get into that now; read her highly informed work on that matter.
This is a serious issue when you realize that publishers are looking at AI-generated art as a cheap and limitless resource. Sci-fi publisher Tor claimed it didn’t know the stock image it purchased was created by AI, though evidence shows that’s kind of naive. James Allen dominated at the Colorado State Fair with his Midjourney-generated artwork, “Théâtre D’opéra Spatial.” The category was digital art, sure, but the other artists still used styluses and their own hands to create their work, while Allen was proficient with a set of text cues.
With all that said, let’s step back and look at the process of “cocreating” AI-generated art. Some programs use reservoirs of creative material that pull from everything the internet can provide; others ethically claim not to grab entire online catalogs without restraint or observe the objection that some artists have over their work being appropriated. DeviantArt, the enormous (and sometimes controversial) online artists’ forum, has provisions in its TOS that claim to protect artists’ work but function like the Honor System: DA might use your work to advertise itself; users may not steal other users’ work—though they do, repeatedly and constantly. One day, DA announced that all artwork uploaded to the site was subject to exploitation by AI art engines, unless artists went through their catalog and opted out of this feature. For some artists, that meant thousands of works stored over several years! Fortunately, DA listened to the explosion of protest and, within 48 hours, reworded their TOS to default to opting out of the program, giving artists the option of opting in if they cared to do so.
Sorry, I keep getting off track. I’m just trying to explain what makes some AI generators better than others. Giving a program access to all possible artwork would produce the best possible creative content; a sparse diet of limited artwork is one factor for creating crappy, unattractive artwork that lacks insight. Another factor would be a crappy program to start with. But after that, it comes down to the cocreator’s skill with words, honestly. The cocreator has to describe a scene that clearly outlines everything they want yet won’t confuse the program, which has limited comprehension skills. Skilled cocreators get familiar with their program-of-choice and can anticipate the gaps in its understanding. As well, once an image is generated, the cocreator usually has an opportunity to massage the image, working variations on it, giving more or less weight to text instructions or granting more or less liberty to the machine to do its magic.
To illustrate this, I fed only the most basic instructions into a variety of AI art programs to see how they would stand up against each other. By keeping it simple, I hope to impart a sense of which programs were adept at interpreting instructions and how broad their resources were. For the most part, all I entered was “cowgirl.” I wanted to see all these AI programs create an image of a cowgirl. When more instructions were absolutely required, I included “a photo of a” or “detailed, realistic.”
Warning: As powerful as it is, AI-generated art still struggles with some basics, like facial structure and hands. You’ll quickly see that AI just doesn’t know what to do with hands, and the uncanny valley quickly fills up with biological anomalies.
BlueWillow: It looks like it started with making a photograph, but the details of the face are more like an acrylic painting, I think. That’s definitely a picture of a woman, and she’s definitely standing in broad plains under big sky. The silver-trimmed belts are a nice touch. BlueWillow is a relatively new contender that’s offering cash prizes to users who recruit the most users, which suggests to me they’re aggressively crowd-sourcing training for their AI program. Will they close their doors to the public and require memberships, once they get strong enough, or will they sell out to some corporation?
Craiyon: Craiyon’s work consistently makes me feel sympathetic for this program. All the other AI programs show clear artistic aptitude, and Craiyon’s struggling to do something it wasn’t meant to do, but it’s really trying its hardest. It knows the general shape of a person, and it seems to have heard of neckerchiefs and vests, though it’s not sure how they go. Still, those are blurry prairie grasses in the background, so good job. Craiyon was one of those programs that tried to ride the wave of AI art created by DALL-E: they called themselves DALL-E Mini and got slapped with a cease-and-desist. But if you’ve seen DALL-E’s artwork… there’s just no comparison.
DeepAI: Full marks on everything for DeepAI. Not only did they put a woman on a horse, they rendered her in silhouette as her mount walks through lavish brush and greens, with an atmospheric haze that really sets the tone. I didn’t ask for trees or even a setting, but DeepAI’s reservoir of knowledge judged these as appropriate to include, and I have no argument. This actually looks like something a talented physical artist would produce, and I can’t spot any egregious errors (other than an extra tail on the horse).
Dream by WOMBO: Yeah, here, you see what I mean about hands. Dream went straight for the photographic effect, which is really of good quality, and something in its database of art resources told it what a cowperson’s belt is supposed to look like. Excellent detail on that bling. The subject is even leaning against a post, kinda, like in a high school photo. Stuff like that can tell you where the program is pulling from, what kind of imagery it has access to. But that hand, damn. We’ve got a long way to go.
DreamStudio: I wasn’t given the opportunity to specify “photo” or “painting” with this one, but DreamStudio produced four paintings and sketches of cowgirls rather than realistic imagery. That’s fine: if I’d run another iteration, I could come away with photographs or woodblock prints or anime, who knows. Again with the standard decorative heavy-leather belt, plus a form-fitting leather vest. The necklace is a nice touch. That hat is running out of control, though.
ERNIE-ViLG: You’re taking your chances with ERNIE-ViLG, let me tell you. It’s a Chinese program, so it takes prompts from other languages, translates them into Chinese, and then runs the image generation. Was the translation accurate? It’s impossible to tell until you get your image, which takes quite an amount of time compared to most of these other programs. I’ll note I haven’t had great luck with ERNIE-ViLG in the past, but it did a passable job with “cowgirl,” right down to the jeans jacket and silver earrings. It’s so interesting to see which details a program chooses to highlight, or maybe that’s all it has access to.
NightCafe: NightCafe is one of the better programs, for my money. I paid for a month’s worth of credits and taught myself the rudiments of text prompts and refinement through this engine (do the kids call these AI “engines”? Just realized I could be embarrassing myself), and it offers the opportunity to run variants on any image it produces, which is very handy when you’ve got a great picture but it’s not quite right. Here we see a decidedly sultry cowpoke, armed with six-shooters unlike any other sample here, and either she’s exceptionally tall or she’s breeding Icelandic horses. She features silver vambraces, for some reason, and you’ll note that her face isn’t quite composed, but if she’s riding beneath a burning sun, those are the least of her problems.
Photosonic: To clarify: Writesonic is an AI writing program (yet another cause for concern for creative types), and it has built an AI art program called Photosonic. This isn’t uncommon: NovelAI and Jasper also started as writing programs that added art capability. Photosonic did a bang-up job with this one: a very clear, coherent black-and-white photo of someone who looks like she’s spent some time on the range. The left leg gets a little messy, and I don’t know what’s going on with those leather straps, but the rumpled denim shirt and proud tilt of the chin persuade me to shut the hell up. Also, I think no other program has rendered so perfect a hat.
Stable Diffusion: This program is offered by HuggingFace, which is an umbrella site with several AI programs geared toward various specialties. ERNIE-ViLG lives here (probably in the corner, playing with its feet), and Stable Diffusion 2.1 is a very strong program in this arsenal. If you weren’t paying attention, you could easily mistake this image for a legit photograph of a Western woman. After a few moments it’s easy to pick out the flaws, but I think it’s more worthwhile to note what it does well. The sheen on the worn leather belt, for one thing, the perfect smile and the sunlight glowing in flaxen tresses. In fact, the sun is clearly shining on only one side of her, so the program is smart enough to have those physics down, as well as placing her in a geographically likely setting.
Through this simple test, I hope I’ve provided a reasonable understanding of the state of this technology. There are many other programs, and some of the best are only accessible for a monthly fee. Some have distinct specialties, like anime/manga imagery, and most (but not all) have NSFW locks imposed: DreamStudio produces blurry images when the results are too salacious, and NightCafe simply won’t accept suggestive or inappropriate terms.
Skilled wordsmiths, using a robust AI program, can produce astonishing work that would take a physical artist half a lifetime to master, no question of that. On the other hand, these programs aren’t coming up with original work. They require a cocreator to feed them the terms, after a programmer has taught it fundamental grammar, and it has to pull from work that’s already been done and identify, to the best of its ability, the desired elements and then arrange them in a way it’s been told makes sense. There’s tremendous potential for error and noise, and when a program comes up against a term it doesn’t know, you have no idea what you’re going to get and there’s no way to remedy that.
Are artists under threat from AI programs? Yes and no. AI can do large amounts of complex work very swiftly, but it cannot innovate, not yet. It’s worth watching AI closely, whether to enjoy it or be critical, but to rule it out as “cheating” or “dishonest” is too simplistic; to banish it on principle is immature. It’s capable of amazing things, but it’s up to us to prompt those amazing things, just like it’s up to us to design the ethics around its use.