A book cover is judged at about 100 pixels tall, in a grid, next to the covers it competes with. That single fact settles most decisions: one focal image, enormous readable type, and a look that signals the genre before anyone reads the title. Generated art is genuinely good at the image half of that. The type half belongs in your editor, not in the model.
TL;DR
- Generate the art, set the type yourself. Models still garble words, and a cover with a typo is fatal
- Cheapest good route: Seedream 5 at 6 credits per image, so twenty directions cost 120 credits, about $1.20
- Genre signalling beats originality. A cover that reads as literary fiction sells literary fiction
- Check at 100 pixels tall in a grid of real competitor covers before anything else
- Two real constraints: print resolution (upscale at 13 credits) and knowing what rights you have
The workflow
- Collect twenty covers from the top of your exact subcategory. Not books you like, books that rank where you want to rank. Note what they share: palette, type weight, whether there is a figure, how much negative space.
- Generate the art only. No title, no author name in the prompt. Describe the image, leave the top third or bottom third visually quiet for type. Seedream 5 at 6 credits.
- Generate twenty candidates, 120 credits total. This is the step people skip and it is the step that works.
- Set the type in your editor with a real font. Your title, your name, correct kerning, editable later.
- Test at thumbnail size in a mock grid with the twenty competitor covers. The one that survives is the cover.
- Upscale the winner for print, 13 credits, and check the bleed and spine requirements of your printer.
Model choice and cost
Canon 8frame prices, a credit is $0.01 at pack rate:
| Task | Route | Credits |
|---|---|---|
| Cover art candidate | Seedream 5 | 6 |
| Sharper tier for the final | Seedream 5 Pro | 7 |
| Cheapest volume exploration | Imagen 4 Fast | 4 |
| Highest-fidelity single image | Imagen 4 Ultra | 12 |
| Stylised, illustrative, non-photoreal | Reve | 34 |
| Art with correct text inside it | Nano Banana Pro | 21 |
| Adjust an existing candidate | Flux Edit (FLUX.2 pro) | 7 |
| Upscale for print | image upscale | 13 |
| Print-grade upscale, per 24MP block | Topaz | 11 |
A complete cover: twenty candidates at 6 (120), two edits at 7 (14), one upscale (13). About 147 credits, roughly $1.47. Inside the $19 Starter plan with most of the month intact, which is the honest comparison against a commissioned cover.
Why the type goes in your editor
Most image models still produce letterforms that look like writing without being it: wrong words, invented characters, spelling that changes between attempts. A book cover is the worst possible place for that, because a misspelled title is a dead listing.
Setting type yourself also gives you what a cover actually needs: exact font choice (which itself signals genre), control over how large the title is relative to the art, and a file you can revise when you change the subtitle.
If you genuinely need text integrated into the scene, painted on a sign or carved into stone within the image, Nano Banana Pro at 21 credits is the model that handles it. Generate three and check the spelling by reading it from the image, not from your prompt.
Genre signalling, which matters more than beauty
Readers browse by pattern recognition. A cover's first job is to say "I am the kind of book you are looking for," and it says that through convention: thriller covers look like thriller covers for a reason that is commercial rather than lazy.
This is where generated art has a specific failure mode: models default to a generic digital-illustration look that signals nothing, or worse, signals self-published. The fix is to prompt toward the convention of your category explicitly, in the language of how the image is made: photographic and desaturated, hand-painted, high-contrast graphic, flat vector, textured paper.
The rights question, stated honestly
Two things worth knowing and not guessing about:
Your platform's rules. Retailers and printers have their own disclosure requirements for AI-generated content and they change. Read the current terms for the platform you publish on rather than relying on any article, including this one.
Likeness and style. Do not generate a recognisable real person for a cover, and be careful with prompts naming a living artist's style. The first is a clear no. The second is unsettled enough that a commercial cover is the wrong place to test it.
Neither of these has a clean universal answer, and anyone who gives you one is guessing.
What a designer still does better
A designer solves the whole object: spine, back cover, series consistency, how the type and image interact, and what the book is actually about. Generation solves the front-cover image. For a first book on a tight budget the generated route is genuinely viable. For a series where covers must work as a set and carry a brand, a designer is still the better spend, and you can hand them your generated exploration as a brief.
FAQ
Can I sell a book with an AI-generated cover? Generally yes, subject to your retailer's and printer's current disclosure terms. Check those directly.
Which model is best for book covers? Seedream 5 at 6 credits for exploration and Seedream 5 Pro at 7 for the final in most genres. Reve at 34 when you want something clearly illustrative rather than photographic.
What resolution do I need? Print requirements vary by printer and trim size; get the spec from your printer first, then upscale to it at 13 credits rather than guessing.
How many candidates should I generate? Twenty. It costs $1.20 and it is the difference between a cover and your first idea.
Generate the art, set the type, test at thumbnail size. The 8frame canvas is free and unlimited, and generation is paid from $19/month.