I Struggled to Design My AI-Generated Book Cover
Last time I covered the release work on KDP. There I mentioned in passing that “AI helped with the cover,” but this cover wasn’t something I could leave at that. I struggled with it, and I thought seriously about copyright too. This time, that’s the story.
What kind of cover I wanted
What I first pictured was a cover that didn’t cram in too much.
Letting AI generate images tends to make you want to pile on elements. But at first I wanted the opposite — negative space and subtraction doing the work. What I had in mind was a simple, abstract composition: an arch like a contrail stretching across the sky, and geometric circles fading softly inward.
(Confession up front: this “initial ideal” would get completely overturned at the very end.)
The process went like this:
- Have AI image generation (Nano Banana Pro) create the base image
- Bring it into Canva, crop, position, and adjust the color, then layer on thin-to-medium serif typography to finish it
That was the flow. Writing prompts, I noticed that specifying materials and textures — “lithograph-style,” “monotype,” “airbrush gradation” — got me much closer to the image I wanted than shape instructions like “circle in the center, lines at an angle.” Not diagrammatic precision, but materials, restraint, and negative space. That was my takeaway this time. And this process — “have AI create the base, then finish it myself in Canva” — connects deeply with the copyright discussion I’ll get to later.
The struggle: it took a long time to reach an answer
This composition wasn’t where I started. If anything, I wandered quite a bit.
The first thing I ran into was a strategy problem. When I consulted AI, it told me: “For Narou-style covers, leading with character art is basically the default. Abstract designs work for general literary fiction, but in the Narou context they’re basically a disadvantage.” I see, I thought. So I resolved to run an abstract concept and a character concept in parallel.
The abstract concept rolled out in this order:
- The first “pitch lines × monochrome × geometric” concept → too polished, rejected as “looks like a slide from a presentation deck”
- Trying to add more of a work-of-fiction feel, I moved toward risograph, lithograph, monotype, ink-wash
- A “duplicated ghost figure” concept hinting at “another life = reincarnation” → rejected for being too diagrammatic
- Following the design of one book cover I used as a reference (a single bold diagonal line on a white ground), I moved toward a concept that let a single brushstroke do the talking
- But that ended up too Eastern, too Japanese-looking. So I mixed in Bauhaus, Swiss design, and airbrush gradation to pull it back toward Western abstraction…
- Finally I arrived at an abstract concept I was happy with: “an arch like a contrail, plus a soft circle”
The character concept had its own separate struggles. The composition was “James seen from behind, standing beside the dugout, gazing out at an empty pitch, an English stadium from the 1980s.” At first I made it in a somber, film-like realism, but it came out too melancholic, too heavy. From there I redid it repeatedly — shifting to cel-shaded anime style, adjusting the downcast pose toward the bearing of a “quiet hero.”
Quietly the most annoying problem was that the generation tool kept inserting an unremovable watermark in the bottom right corner. Prompts couldn’t get rid of it, so I resorted to little tricks like composing the image so that no important elements sat in that corner to begin with.
Running through all of this was a single conviction: I didn’t want an image that screamed “AI made this” at first glance. Rather than aim for perfection in one shot, I generated many versions and picked from them. I didn’t want to make the face of a book look cheap — that tension was what ate up the most time.
In the end, one advertisement I saw on the street flipped everything over
The abstract concept wasn’t bad. I liked it myself. …Or so I thought.
Then one day, I saw an advertisement for a Narou-style work on the street. Every single work lined up there had its character front and center, looming large. In that instant, it clicked into place. “Ah, right — in this genre, you have to put the character front and center.”
I had understood in my head what AI told me at the start — that leading with character art is the default for Narou — but somewhere in my heart I’d been prioritizing my own aesthetic sense, wanting to nail something cool and abstract. But what matters for a cover meant to get someone to pick up the book isn’t my aesthetic sense. It’s the reader instantly recognizing, “this is the kind of thing I like.” That street advertisement drove the point home.
So I dropped the abstract concept outright and made a full pivot to a manga-style with the character front and center. What came out of that is the current cover — the protagonist James, sitting on a bench, gazing at a pitch after the rain. Behind him, a 1980s stadium; beside him, the rest of the bench; at his feet, a tactics board. The title, large and hard-edged: “EXTRA TIME.” After all that wandering through abstraction, the “character concept” I’d started with ended up snatching the leading role at the very end. Ha.
(This is the cover version of the “I went along with the Narou template” story I wrote about last time. Here too, in the end, the right answer was to obediently follow the conventions of the genre.)
Since I was using an AI-generated image for the cover, this was something I couldn’t avoid. I’m not an expert, so I can’t state anything definitively, but I’ll leave a record of the points I checked and considered on my own.
- How copyright treats AI-generated works: Whether copyright arises in the generated work itself, and if so, who it belongs to. As with the main text, I understand this is an area where the degree of human creative involvement matters (the general framework being that images lacking human creative contribution are less likely to be protected).
- Risk of resembling existing works: Whether, due to its training data, the image ends up resembling a particular artist, work, or character. This work in particular is set in 1980s English football and features real club names in the story, so I was warned that if the cover ended up incorporating elements close to a real club’s emblem, colors, or a real player’s likeness, that would be dangerous. So even after deciding to put the character front and center, what I drew was strictly my own original character (James) and a fictional club (Wandle). I was careful not to resemble any real club or player, and to make the emblem fictional as well.
- Terms of use for the tools, and whether commercial use is allowed: Whether the generation tool I used, Canva, and the fonts were licensed for use in commercial publishing. I checked this including the license for the serif font. As the cleanest option from a rights standpoint, I also considered something like Adobe Firefly, whose training data is licensed and comes with indemnification.
- KDP/Amazon’s own policies: The platform’s content requirements. KDP draws a line where “AI-generated” content requires disclosure, while “AI-assisted” content does not, so I checked which category my process fell under.
- How to disclose AI use: Including for the main text, how much and in what way to make this explicit.
In the end, the design swung dramatically from abstract to character-front-and-center. Even so, my basic policy never changed: use AI to generate “material,” but make the final compositional decisions — cropping, placement, typography — myself.
To put it concretely — I have Nano Banana Pro create the base image, but I don’t stop there. I bring it into Canva and go through it myself, again and again, deciding what to crop, how to lay the text, how to adjust the colors, until it becomes a finished “cover.” I came to think that it’s precisely in this editing process that human creative judgment sits. Using a generated image as-is and taking it as material to recompose yourself are utterly different in the degree of “human creative contribution” — and that, I think, is what matters both for thinking about copyright and for thinking about KDP’s line between “AI-generated” and “AI-assisted.”
Curiously, this is exactly the same philosophy as how I made the main text — let AI write, but keep the final judgment in human hands. For both the cover and the text, how much human creative contribution I could preserve was my own guiding measure.
Just to be clear — I’m not a lawyer, so nothing written here is a legal conclusion. This is an area where the rules are still in motion, so I hope it’s read as a record of “here’s how I looked into it, and here’s what I decided,” which might be useful as a reference for others putting out work in the future. (Please check the actual terms and licenses for yourself, using the most current versions.)
Next time, we’ll get back to the substance. I’ll write about the one “heavy theme” I put into this story that I started with such a light touch.
Below are the remains of the failures
Originally published in Japanese at https://clazytech.com/2026/06/1654/. Translated with LLM assistance and reviewed before publication.