One project I’ve worked on since submitting the first book to the publisher (other than writing the second book) was trying to find concept art for the cover. I’ve had quite a few people ask on my social media where this came from:
The answer is MidJourney image generation.
AI Image Generation
As most of you have probably heard, there’s been a tremendous ruckus lately about so-called artificial intelligence and what it’s now capable of. If you’re not aware of the discussion, this is the most interesting analysis I’ve found to date on what AI is right now and what it could be capable of. (While we’re at it, here is another fascinating discussion on “evil” AI I found recently.)
Regardless, current image generation is not based on that same model, but it is capable of creating images from scratch based on a prompt.
Last I checked, MidJourney limited their service to those who are willing to pay. For my purposes, even if the images only serve as AI concept art, it was worth at least a month’s basic subscription to create concept art for a book cover.
How it works
First, you message the MidJourney AI bot with a prompt. Mine was something like the following:
Landscape of a peak overlooking billowing mist under a night sky full of stars with colliding galaxies. Profile view of a futuristic, teenage boy with a crew cut, wearing a white shirt and white pants with a white belt, stares to his right at a particularly bright star across the image, 8k, UHD, hyper-realistic, ultra detailed, intricate details, cinematic lighting, full hd render
Then, the bot generates a sample of four images. It looks like this:
From there, you can choose an image you like and either upscale it or create variations. This set of images was a variation on the second option above:
From here, you can choose a final version to upscale or continue creating variations. I chose option three and upscaled it to get the image at the top of this post (this one):
Some Observations
MidJourney’s algorithm isn’t very good at following specific instructions to start out, and it doesn’t seem to learn over time as fast as I expected. I’m not sure whether this was user error or just limitations of the system.
For example, I gave specific instructions in the prompt about clothing color and the positioning of the subject. It almost never got both factors correct.
Specifying the location of the the boy made no difference. Sometimes he was to one side as I requested, but usually he was centered in the shot. (My prompt was more specific on that point in some of the early attempts.) I couldn’t tell whether this improved as I iterated or not.
My conclusion
I got a lot of results I liked in different ways. Even with hundreds of attempts, though, I couldn’t get them all to combine into something I really loved. (My original aim was a boy, looking from one side of the image to the other at a star, on a desolate peak overlooking clouds with a galaxy showing across the sky.)
I also found that the style and stylization of the image varied considerably.
This was one of the early results that I particularly liked.
And this was one of the later ones that was closer to what I envisioned, but in which the art style was less pleasing to me:
Because I haven’t received any response from this publisher as yet, I don’t know whether I’ll end up hiring an artist myself or using their inhouse art department. Either way, AI concept art seems promising. I’ll DEFINITELY be using MidJourney to create a title illustration for things like short stories!
