This post is dedicated to Janelle Shane, who sparked my interest in AI learning.
Now we’ll move out of the realm of the human artist and look at what art-creating AI engines can come up with. These engines work with a text prompt like “Fox in a field of flowers” or “Volcano erupting over Los Angeles.” For my prompt I used “Evil queen of an ancient empire with long black hair.” Because that, to me, is the essence of what Jadis is.
This first rendition is from Nightcafe.com.
This looks like Hawaii to me, perhaps an interpretation of Honolulu’s towers reflected in a quiet ocean. Something is burning (Charn?) and Jadis is regal in a purple cape, though she also appears to be a cyclops, and is melting into a pair of gray stone buttocks.
This one is from is from Craiyon.com, formerly Dall-E (Salvador Dali, get it?) Mini.
Craiyon generates nine different scenes from the same text prompt, arranged Brady-Bunch style. The one above was the most interesting of the first batch. The render is rough, but compositionally it’s good. Jadis has long black hair and red slits in her black velvet gown, and she poses defiantly leaning on her right elbow while her body makes a three-quarter turn. Her face looks evil yet glamorous. Unfortunately she also has an extra set of hands, one at her left hip while the other appears to be holding a dagger, which wasn’t in the text but an element the AI filled in on its own!
The more specific prompt “Jadis, Queen of Charn” generated the set of nine below, in wildly differing styles as all the artworks with those words were searched and blended.
The top center and center ones are pretty good for an AI.
A search a day later produced even finer creations as the AI engine “learned” to satisfy the same prompt. These two were the best.
The left one looks almost perfect — that is, created by an actual human — save for some wonkiness with the Queen’s eyes and jawline and the vague tangle of goldwork that is her crown. Likewise for the one on the right, in which she has a black eye, an abrasion near the corner of her mouth, and blood trickling from her right nostril for some reason. Perhaps because she was “beaten” by Aslan?
Craiyon has a forum so users can share their often hilarious creations. The site is a work in progress and is likely to continue changing in the future.
Jadis as interpreted by Starry AI. You’ll need a Google or Apple account to use it online, and it’s also available as an app. This site gives you a beginning choice of the Altair or Orion engines, the first dreamlike and abstract, the second “unreal reality” in the engine’s own words. In addition the app offers 37 artistic styles ranging from Rococo to Thomas Kinkade. I chose “National Geographic” for my girl. The art is generated in multiple layers, so you can actually witness the AI’s visual mind chugging away.
So, in this awesome pic we have an ancient empire, in ruins and burning as appropriate for Charn. But Jadis herself is growing out of a rock. Her face is smudged from the fire and quite pudgy, and she is wall-eyed with either a fat lower lip, or is sticking her tongue out. Well, they say great evil is reflected, eventually, on the face.
But what’s interesting is that the AI interpreted “with long black hair” as a flowing mane of long black hair that is actually standing beside her like a companion. This too is growing out of rock, has its own golden crown, and is turning sulphur-yellow at the tips. Note also the random non-English characters to the left of Jadis’s head, which seem to be explaining the scene as if in a magazine.
(An important key to using all of these sites is to consider what the original images might be titled and what words might be in that title — which is also a basic principle of SEO optimization. For example, it’s more likely a picture of a lion would include the words “Lion” or “Felis Leo” than the phrase “a large golden-brown cat with a mane living in Africa” which is something a crossword puzzle writer would use.)
Dream by Wombo works a lot like Starry AI does, giving the user a choice of styles and also requiring a Google or Apple account. “Jadis, Queen of Charn” using the “Fantasy Art” style netted the following image.
The AI did Charn correctly, to my eyes: a never-ending city towering into the sky and climbing down into the valleys, though the style is more Pablo Picasso in his Cubist period than Michael Whelan. But Jadis, even giving she has giant blood, is more than giant-sized… she’s kaiju-sized, towering over her city in a flowing gown. She’s also bald and looks more than a little geriatric. Perhaps as the AI learns it will get the picture more correct.
Replicate.com offers art generation for the geekiest of the geeks. It’s not as user-friendly as Nightcafe or Starry AI, but it does offer a bevy of models, allows download of the programs, and lets you write your own coding, if you’re inclined. One of the generating models is Erlich, which generates “a logo with text” though in truth it can be any kind of graphic image with text, such as a book cover or cereal box. Here’s four examples of art from the prompt “Jadis, Queen of Charn.”
Instead of a logo, here we have what might be rough renditions — thumbnails, we call them in graphic design world — of book covers or box lids of role-playing games. Erlich did a smash-up job with the exotic costumes of the top pics. The one on the left is a hybrid of Tibetan priest and goldfish, and the right, Samurai and Scythian warrior. Erlich had problems with the name Jadis, however. The -is at the end is correct, but not the d in the middle, and in the lower left pic, the hook of the J.
The real surprise, however, is the pic at the lower right. It’s been modeled on a MAGIC: The Gathering card!
Another engine on the Replicate site is Pixray, an image generator. It created these images from the text prompt “Jadis, Queen of Charn” using two different art styles.
The woman at the top resembles a nightmarish caricature of Anna Wintour, while the second makes me think of Ursula K. LeGuin with an oversized head.
I don’t think AI will replace human artists any time soon. But for generating thumbnails and rough ideas, they are already there.