If you recall, last April I posted a couple of entries about my experiments with AI image generators. There were serious problems drawing hands, feet, and faces. The other day I got an email saying that the Gab social network had installed an AI image generator called Gabby that registered users could try for free. So I tried it.
I have two general test categories of images I would like an AI to generate: Pictures of a thingmaker from my drumlins stories like “Drumlin Boiler,” and pictures of a woman sitting in a magical basket flying over downtown Baltimore, from my still-unpublished novella, Volare! I tried them both, and will include the best images from my tests below.
The drumlin thingmaker is a relatively simple structure: a 2-meter-wide shallow bowl made of what looks like black granite, half-full of a silvery dust, with two waist-high pillars in front of it, one smooth, the other vertically ridged like a saguaro cactus. In the stories, people tap a total of 256 times on the tops of the pillars in any combination, and the machine will then build something in the bowl. There are 2256 different possible codes, in base 10 1.15 x 1077, which is in the vicinity of the number of atoms in the observable universe. The people marooned on the planet where the thingmakers were found learn to use them, and I have several stories about the alien machines and their products, which thingmaker users call “drumlins.” (I know a drumlin is a glacial landform. I’ve repurposed the word, as SF writers sometimes do.)
As with the other image generators, you begin with a statement of what should be in the image. For the woman in a basket, I used the following prompt:
- A barefoot woman in pajamas sitting in a magical wicker basket flying over downtown Baltimore at dawn.
The best image I got was this:
Impressive, compared to my earlier efforts. The woman is African-American, which doesn’t matter; after all, I didn’t specify the woman’s race and Baltimore is a mostly-black city. The basket is wicker. The city does look like Baltimore. (I used to live there in the mid-‘80s.) So far so good. However, on the one foot we can see, she has two big toes. And it took Carol only seconds to note that she has two left hands.
Alas, she isn’t flying but rather sitting on the edge of somebody’s roof. I did specify “flying.” So I give it a B-.
I did a lot better in some ways with the thingmaker. The prompt I used for the image shown below is this:
- A 2-meter wide shallow bowl in a forest clearing, made of polished black granite, half-full of silvery dust, with two polished black granite pillars behind it.
The best image for this test is below:
The bowl is actually pretty close to what I imagine a thingmaker bowl looks like. It should be a little shallower. The two black pillars behind it look like trees. Ok, I didn’t specify how tall the pillars should be. My bad. But the dust is simply missing. I guess I should be glad that it didn’t build me a picture of Oklahoma in the 1930s.
Before I ran out of my daily limit of generated images, I decided to start from scratch with the woman in a basket. In Volare! the basket is a wicker basket about 3 feet in diameter, half-full of weeds that my female lead Edy Gagliano had pulled from her garden. So I began with this prompt:
- A 36" wicker basket half-full of weeds.
How hard could it be? Well, Gabby handed me a wicker basket with plants in it. However, it wasn’t a basket of weeds but a flower arrangement. I tried twice with the same prompt, and got the same thing: live plants in a basket, at least one suitable for putting in your bay window. The weeds were described in the story as wilted dandelions recently yanked and probably wilted if not dead and gone brown. No luck.
In a way I can’t bitch: These are all pleasing images, and Gabby doesn’t have the same problem with plants that it does with hands and feet. And the woman’s hands and feet are mostly better than what I got with Dall-E last April. We’re making progress.
Now, I don’t intend to use an AI-generated image directly as a book cover. There are some weird and currently unsettled copyright issues involved with AI graphics, largely concerned with what content the AI is trained on. I’ve heard rumors that Amazon is yanking self-published books from the Kindle store if it looks like they have AI-generated graphics as covers. That’s an easy enough bullet to duck: I’ll do as I’ve always done and commission a cover from a real live artist. The AI images would be used to suggest to the artist how I imagine various elements of the cover.
This was fun, and if you know of any other AI image generators that you can use without paying for them, please share in the comments, with a sample if you’re so inclined.
I use an image generator called NightCafe for creating background images for Majestic Hearts show flyers. I ran your “thingmaker” test phrase through it and got this. The bowl is shallower, but maybe too shallow, there’s no sign of the pillars, and the “dust” inside the bowl looks more like water.
It’s not a bad image. (I downloaded it; thanks!) Yes, a little on the shallow side, and half full of water rather than dust. But what I will ultimately do is staple a few of the better images together when I ask an artist to draw something specific for me. My story Volare! sat half-finished in a box for almost forty years until early 2023, when I finally finished it. It can wait another year for a cover. I’m trying to finish the first draft of The Everything Machine and it’s no small project.
See this link for the thingmaker image: https://cloud.erbosoft.com/s/YLasP6LkDs6KdZK
The service is at https://creator.nightcafe.studio.
(Tried to use HTML in the comment text, didn’t work…)
I took a close look at the woman’s foot, and the second toe looks OK to me. It’s longer and narrower than the first (“big”) toe; maybe not as much narrower as usual, but enough to be distinct..
As to the third image: you didn’t specify dead or dried weeds.
Her toes still look odd, but the rest of her is pretty striking. She has two eyes and as best I can tell, the right number of digits on her hands and feet. That’s certainly a better result than I had back in April. Two left hands? I didn’t even notice until my wife (who is a retired physical therapist and knows something about human skeletal structure) pointed it out.
As for the weeds, yeah, my bad. Next time it’ll be “a wicker basket half-full of dead dandelions.”
I can definitely say that Amazon KDP *asks* you whether any of the cover or content was AI generated. Very insistently, every time you update. So I imagine bad things are going to happen with that soon.