Castle Falcon was the first book I self-published years ago. At the time, I couldn’t afford cover art from professionals, and generated a cover using the only tools I had: the AutoCAD/AccuRender engineering combination I used to make realistic images of spacecraft in orbit.
I came up with a concept illustrating some of the major features of the book, and either purchased, created, or kludged 3D models of things like a dragon, a myrmidon, a wall, and a stained-glass window. It took a long time, and a lot of post work with Photoshop, but the cover came out okay. This was the wraparound cover I used for a dust jacket hardcover version:
I still dreamed of having a pro do a nicer cover. I had imagined one layout showing the main characters Katie and Zach exploring the outdoor grounds of their huge castle home. Katie would be running down an ancient set of stairs, while Zach would be running his skateboard down the wide stone rail. In the background would be the countless esoteric and ancient structures of Castle Falcon, and in the distant haze the Black Tower would rise above all the lower buildings with the Dragon perched atop it. It would be perfect if Michael Whelan did it. Ah, well. Like most pro artists I’d ever approached, I never heard back from him. Hey, I at least wanted to know exactly what I couldn’t afford, right?
And then came AI imaging. I described how I redid my cover for Zorya, which turned out surprisingly good for surprisingly little effort. Could my dream cover be realized the same way?
I made a few attempts at a long and complex prompt that would spit the final image out, and as you might expect, got hash in return. One element might be good, but the others weren’t. After a while, I realized I’d never get what I wanted that way. So I decided to construct the elements separately, and Photoshop them together.
Start with the background. I created this prompt:
Ground view of large area of old buildings of many historical styles. Ivy grows on some. Grass surrounds them. In the distance, a single tall black tower rises. In the foreground is an ancient set of stone steps with railings.
You will note that I did not say “in the style of Michael Whelan.” I avoided this since it seemed to me it would be an engraved invitation to encourage the program to harvest his work. I later learned that the AI would have specifically rejected it anyway, but it seemed a good precaution.
It took only a few generations to give me this:

This was amazingly close to what I had in mind, although in my imagination the stairs ran from upper left to lower right. I could work with this. The Black Tower was all wrong. It was supposed to be a thin tapering spire with a chamber at the top. I wasted a lot of time trying to get the image to change that, or add the two kids, just to see if I could do it, but no dice. I just used normal image editing to “delete” the tower, and went on from there. I opened a Photoshop .psd file and went to work.
Start with Katie. The prompt: Girl running down steps, 12 years old, blonde, t-shirt, jeans shorts, and tennis shoes.
Again, it took only a few attempts to get a workable shot. Several versions insisted on having her run up the stairs.

I flipped it horizontally so she’d be running the right direction, and went on to Zach. The prompt: Distant shot, 11 year old blonde boy, dressed in t-shirt and shorts, with bare feet, riding a skateboard down a ramp. Hair is cut short and spiky.
Most of the attempts had him going the wrong way entirely, like moving away from me. But this one was perfect.

I used up some time trying to describe the Black Tower effectively in a prompt. I won’t list the prompts since I didn’t use them anyway. I wasted even more time in a dead end trying to get a tower with a dragon on it. What I finally did was dive into my Solidworks program and just model and render the very simple Tower design:

There we go. Finally, I needed to perch a separate dragon on top of the tower, so I had to generate a dragon. That took more time than any other element. AI has trouble with dragons. Here’s an early Dall-E attempt with an older prompt. The program had the same trouble with dragon wings as it did with hands. I switched back to Design and kept going.
The final prompt: winged dragon, gunmetal color, perched on a steeple clear sunny sky. I figured I could just delete the steeple if I got the right pose. There were many dragons, some of them fairly decent, but this one had the look I wanted:

Clipped it out from the background, flipped it horizontally, and then went to work in Photoshop. I whipped up layers and pasted in the kids, the tower, and the dragon. Then came the tedious process of making them look like they weren’t pasted in. Borders needed a bit of blur to match backgrounds. Shadows had to be created for the kids, and some goofy AI stuff like the railings in the foreground of the background shot had to be tweaked. Katie’s proportions needed a little work. I had to put a hint of electric blue in the children’s eyes. The tower and dragon had to have some distance put on them. And then I had to use an AI image extender to increase the height of the entire image so there’d be plenty of room for a title if I used it for a cover.
This was the end product:

Detail of the kids:
I haven’t done anything to make a final book cover out of this yet. We’ll see what happens. I notice that there’s a lot of “matte painting” style to large AI landscapes. Detail suggests itself to the eye even if it tends to break down under magnification. You see what you expect to see. A real painting examined very closely does this too.