AI Fantasy Book Cover Prompts: Composition, KDP Disclosure Rules, and What You Can't Copyright

How does a book cover prompt differ from a normal character portrait prompt?

The character description doesn't change — a good cover prompt still names ancestry, permanent visual anchors, gear, and mood the same way any portrait prompt does (the prompt composer output works as-is for this part). Four things change around it.

Aspect ratio comes first, not last. Generate portrait-oriented from the start. Cropping a square image to 1.6:1 throws away composition; the model composes differently when it knows the canvas is tall. In Midjourney that means --ar 2:3, or --ar 5:8 if you want to match Kindle's 1600x2560 spec exactly.

The subject shrinks. A portrait fills the frame with the face. A cover usually places the character in the lower two-thirds — half-body or full-body framing with room above. "Character positioned in the lower third of the frame" is a phrase image models follow surprisingly well.

Thumbnail readability beats detail. Amazon search results render your cover at roughly 100 pixels wide. One strong silhouette, high value contrast between subject and background, and a two-to-three-color palette survive that. Intricate armor filigree and busy backgrounds turn to mud. Squint-test every candidate at 10% zoom before you commit.

Mood carries genre. Browsers decide genre from color and light before they read a single word. A cold, grim palette says grimdark; golden, warm tones say classic epic fantasy. Pick the palette for the shelf you want to sit on, not the one that flatters the character.

How do you leave clean space for the title and author name?

You need two quiet zones: the upper quarter to third for the title, and a strip near the bottom for the author name. Three prompt techniques stack to produce them.

Describe the empty area as content. Models fill space unless you tell them what the emptiness is made of. "Vast pale storm sky fills the upper third of the frame" works better than "empty space at the top." Other reliable fillers: "background fading to near-black above the figure," "soft fog dissolving into darkness," "smooth gradient of deep twilight blue."

Say "negative space" anyway. The phrase appears in enough design-annotated training data that it nudges composition toward minimalism. Combine it with placement: "generous negative space above the subject."

Ban text generation. Add --no text, letters, watermark in Midjourney, or "no text or lettering" in prose for DALL-E. AI-rendered lettering is mangled often enough that you should always set real type in Canva, Affinity, or Photoshop with a properly licensed font.

A weathered human ranger in a hooded oak-green cloak stands in the lower third of the frame, full body, seen from a low angle looking toward distant snow-capped peaks. A vast pale storm sky fills the upper two-thirds of the frame, soft unbroken cloud cover, generous negative space above the figure. Muted, desaturated palette of slate blue, ash gray, and cold moss green. Painterly digital illustration, epic fantasy book cover composition, strong readable silhouette. --ar 5:8 --no text, letters, watermark

If a generation you love is too tight, don't reroll — extend it. Midjourney's editor can zoom out and pan upward to add sky, and Photoshop's Generative Expand does the same job on any image. Adding headroom after the fact is far cheaper than chasing a new composition.

What sizes do KDP ebooks and ACX audiobooks require?

Numbers straight from the platform specs, current as of mid-2026:

Kindle ebook (KDP). Ideal size is 2,560 px tall by 1,600 px wide — a 1.6:1 height-to-width ratio, which reduces to exactly 5:8. Minimum is 1,000 x 625 px; anything under 500 px on the shortest side won't display on the site. JPEG or TIFF, RGB color (no CMYK), under 50 MB. Amazon compresses whatever you upload, so upload lightly compressed.

Paperback and hardcover (KDP). Print covers are a single full-wrap PDF — back cover, spine, front cover — whose dimensions depend on trim size and page count. Use KDP's cover size calculator to generate the exact template with spine width and 0.125-inch bleed, and build the wrap at 300 DPI in a design tool. The AI image becomes the front-cover panel, not the whole file.

Audiobook (ACX/Audible). A true square, no smaller than 2400 x 2400 px, RGB, in JPG, PNG, or TIF — and the cover image itself must include the title and author name. A 2:3 ebook composition doesn't crop to square gracefully, so either generate a separate --ar 1:1 variant of the same scene or extend the canvas sideways with outpainting.

On raw output size: ChatGPT's image tool tops out at 1024x1536 in portrait, and Midjourney's default renders also land below Kindle's ideal spec, so plan an upscale pass (the in-tool upscaler, or a dedicated one like Topaz) and confirm you clear 2,560 px on the long side before upload.

What does Amazon make you disclose about AI cover art?

KDP's content guidelines draw one load-bearing distinction.

AI-generated means "text, images, or translations created by an AI-based tool." If Midjourney, DALL-E, Stable Diffusion, Flux, or Leonardo produced the artwork, it's AI-generated — and KDP is explicit that it stays AI-generated even if you edit it substantially afterward. You must disclose it during publishing, and the requirement covers cover art and interior images, not just text.

AI-assisted means you created the content yourself and used AI tools to edit, refine, or error-check it — or used AI only to brainstorm while ultimately creating the artwork yourself. AI-assisted content does not require disclosure.

A prompt-generated fantasy cover is unambiguously in the first bucket. Retouching it in Photoshop, color-grading it, even repainting sections doesn't move it to the second.

Three practical notes. First, the disclosure is made to Amazon in the publishing flow; as of mid-2026 it is not displayed on your product page, so checking the box costs you nothing with readers. Second, failing to disclose is a content-guidelines violation — Amazon can block or remove the book, and repeat violations can end the account. Given how cheap compliance is, gambling is irrational. Third, disclosure doesn't shift responsibility: you still warrant the image doesn't infringe anyone's IP, so prompts that name living artists or trademarked properties are a separate risk you carry, not Amazon.

If you commission a human artist who uses your AI generations only as reference and paints an original cover, the artwork is human-created and lands in AI-assisted territory — that's the one clean route from bucket one to bucket two.

Can you copyright an AI-generated cover, and what makes it protectable?

Under current US law: not the raw image. The Copyright Office's January 2025 report Copyright and Artificial Intelligence, Part 2: Copyrightability concluded that prompts alone don't give a user enough control over the output to count as authorship — no matter how detailed the prompt or how many times you rerolled. This wasn't new; the 2023 Zarya of the Dawn decision had already registered a comic's text and its arrangement of images while refusing the individual Midjourney images themselves.

The practical consequence for authors is blunt: the raw generated image on your cover is effectively public domain in the US. A competitor who grabs it and puts it on their own book isn't infringing your copyright, because there isn't one. (Midjourney's terms make paid subscribers the owner of their assets contractually — and companies over $1M in annual revenue need a Pro or Mega plan for that — but a contract with Midjourney doesn't create copyright against third parties.)

What is protectable is your human contribution:

• The cover as a whole. Your typography, layout, and the selection and arrangement of elements are human authorship, exactly like Zarya's protected arrangement.
• Substantial creative modification. In January 2025 the Office registered A Single Piece of American Cheese, built in Invoke through dozens of documented inpainting selections and arrangements — the human choices were the protected authorship, not the AI pixels.
• Documentation. That registration leaned on a process video. Keep your working files and iteration history.

So composite, overpaint, and design — don't just export. None of this is legal advice; for a flagship series cover, an hour with an IP attorney is cheap.

What do example fantasy cover prompts look like by genre mood?

The recipe: compose the character in the generator, then swap the portrait framing for cover framing and bolt on the genre's palette, lighting, and the --ar and --no parameters. Three worked moods:

Epic fantasy — small figure, vast world, warm golden light, warm palette. The ranger prompt above is the template: full body low in the frame, sky as title space.

Dark romantasy — intimate framing, jewel tones against near-black, one dramatic light source. Rim light is the workhorse here because it carves the silhouette while leaving the background empty for type.

A pale drow sorceress with silver-white hair and luminous violet eyes, half-body, positioned in the lower left of the frame, turning toward the viewer with a guarded expression. A single ember-orange rim light traces her profile against a background that fades to near-black in the upper half of the frame, clean negative space for typography. Rich saturated jewel tones of amethyst, garnet, and candle gold. Dramatic oil painting style, dark romantasy book cover mood. --ar 5:8 --no text, letters, watermark

Grimdark — desaturate everything, then let one color survive. "Muted ash-and-iron palette, a single blood-red banner as the only saturated element" gives the thumbnail a focal point without brightening the mood; pair it with storm light and a scarred mercenary or necromancer subject.

Whatever the mood, keep one wording discipline: if the cover character also appears in your marketing art or later books in the series, reuse the identity anchors verbatim between prompts — the same drift rules from keeping an AI character consistent apply doubly when the face is your brand.

Frequently asked questions

Does Amazon show the AI disclosure to readers on the product page?

No. As of mid-2026, the AI-content disclosure you make during KDP publishing is internal to Amazon and does not appear on your book's product page or in search results. Readers see nothing different. Amazon could change how it uses the information, but currently disclosing an AI-generated cover has no visible effect on your listing, while failing to disclose risks the book being blocked or removed.

What size does a KDP paperback cover need to be?

Paperback covers are a single full-wrap file covering back cover, spine, and front cover, so there's no one fixed size. Dimensions depend on your trim size and page count. Use KDP's cover size calculator, which outputs a template with the exact spine width and 0.125-inch bleed for your book, then build the wrap at 300 DPI and export as PDF. Your AI image only fills the front panel.

Can I legally sell a book with a Midjourney-generated cover?

Yes. Midjourney's terms of service give paying subscribers ownership of the images they create, including commercial use; companies with over $1,000,000 in annual revenue must be on a Pro or Mega plan for those rights. Selling the book is fine. What you don't get is copyright in the raw image itself under US law, so that ownership is contractual with Midjourney, not enforceable against a stranger who copies the artwork.

If I heavily edit the AI image in Photoshop, does it become copyrightable?

Sometimes, and only for what you added. The US Copyright Office assesses human contribution case by case: creative selection, arrangement, compositing, and substantial repainting can be protected, while the underlying AI-generated material stays unprotected. The 2025 registration of A Single Piece of American Cheese succeeded because the creator documented dozens of human inpainting and arrangement decisions. Keep working files and iteration history as evidence. Note the edits do not remove Amazon's disclosure requirement.

Should I let the AI generate the title text on the cover?

No. Even current models that render legible text mangle letterforms often enough to look amateur at print resolution, and you can't kern, resize, or restyle generated lettering later. Generate the artwork with no text, then set the title and author name in Canva, Affinity Publisher, or Photoshop with a properly licensed font. Hand-set typography is also part of the human authorship that makes the finished cover protectable.

What happens if I don't disclose AI content to KDP and get caught?

Non-disclosure violates KDP's content guidelines. Amazon can block the book from publishing, remove it from sale, and terminate the KDP account for repeated violations, which forfeits the catalog attached to it. Since the disclosure is currently invisible to readers and takes one checkbox during setup, there is no upside to omitting it. If you previously published a book with undisclosed AI content, you can update the disclosure when you next edit the title.

Do I need a different cover image for the audiobook edition?

Usually yes, at least a recomposition. ACX requires a true square at 2400 x 2400 pixels minimum with the title and author name in the image, while a Kindle cover is a tall 1.6:1 rectangle. Cropping a vertical composition to square typically beheads the figure or kills the negative space. Generate a square variant of the same scene with an identical character description, or extend the canvas sideways with outpainting.