How to Make a Consistent Picture Book With AI: Lock the Character, Text, and Print Layout

What the five-generation build actually produced

The fictional test book is Pip and the Paper Moon. Pip is a young red panda with a white cheek-and-brow mask, a mustard hooded raincoat, two navy buttons, teal boots, a cobalt satchel, a striped tail, and a stitched crescent on the anatomical right sleeve. The first generation put the same character in front, three-quarter, side, and back views. That image became the anchor for every story scene.

The four scene prompts repeated the identity block and added only the new action, setting, continuity state, and reserved text area. Each call also received the anchor and the approved preceding scenes that still fitted within the reference limit. No seed was available, and the built-in tool did not expose a model name or model version.

This is a visual editorial review, not an automated identity score. I accepted all four scene generations without a rerun because the locked traits remained usable and the one-button view had a clear pose-based explanation. A missing button on a fully visible coat would have failed the same review.

The locked manuscript contains four lines, one for each scene. I placed each line on its own facing page instead of asking the image model to draw words. The generated images are 1,774 by 887 pixels. At 4.5 inches wide in the 5 by 7 inch no-bleed layout, they retain about 394 pixels per inch. The PDF has a title page, four text-and-image pairs, and a production-note page.

The downloadable PDF, prompt record, and generation record preserve the exact story text, the submitted prompt blocks, file hashes, image dimensions, review observations, and known limits. The ten-page proof cannot be submitted as a KDP paperback because KDP's current minimum is 24 pages. It has not been uploaded to KDP or checked as a physical proof.

Set the physical book before generating finished art

Trim size, page count, bleed, binding, paper, and color option change the space available to every illustration. Record them before making final scenes. Otherwise, a square image may need a destructive portrait crop, a face may land in the gutter, or the selected printing option may reject the planned page count.

KDP provides a useful worked example. Its current paperback guidelines require full-bleed art to extend 0.125 inch (3.2 mm) past the top, bottom, and outside edges. A full-bleed interior PDF page is therefore 0.25 inch taller and 0.125 inch wider than the selected trim. KDP accepts single pages rather than two-page files and requires images of at least 300 DPI.

For an 8.5 by 8.5 inch KDP book with full bleed, each interior PDF page is 8.625 by 8.75 inches. At 300 pixels per inch, the placed artwork needs about 2,588 by 2,625 usable pixels after cropping. The general calculation is:

required pixels = final placed size in inches × required pixels per inch
effective ppi = usable pixels after crop ÷ final placed size in inches

Run the calculation in both directions. A 3,000-pixel-wide image placed at 10 inches is 300 ppi. If its final crop retains only 2,400 pixels across the same printed width, the effective resolution is 240 ppi.

Map the manuscript to pages before illustrating scenes

Create a pagination map with one row per page or spread. Give each row a stable ID and record the exact manuscript text, pictured action, required characters and props, continuity state, and reserved text area.

This map keeps a later text edit from silently changing the illustration brief. Rough each spread with boxes or sketches first. Show the gutter, trim, bleed, safe area, focal subject, and text frame. Plan a spread as one composition, then export it as separate pages when the printer requires single-page files.

Keep eyes, mouths, hands, and story-critical props away from the center fold and trim edges. Background color and texture can cross the gutter. Information the reader must see should remain understandable when the book is held naturally and the pages do not lie flat.

Freeze the character before building the story scenes

Write a character contract with visible traits that can be checked from one scene to the next. Mood words such as "cute," "whimsical," or "cinematic" may guide style, but they do not establish identity.

CHARACTER LOCK
Age or species:
Body proportions and height relative to other characters:
Face or muzzle shape:
Hair, fur, or skin color:
Eyes and eyebrows:
Outfit colors and construction:
Shoes and fixed accessories:
One-sided details, described from the character's anatomy:
Features that must never appear or disappear:

Generate and approve a reference sheet before the first story spread. Include the views, expressions, and fixed accessories required by the manuscript. Use plain lighting and a simple background so shadows and scenery do not look like character traits. Keep the accepted file, prompt, tool or model version, date, and visible review notes together.

Attach that approved reference whenever the image tool supports it. ChatGPT Images accepts an uploaded image and a written edit request, according to OpenAI's current image instructions. Midjourney's current editing documentation describes reference-image use for character consistency.

A seed number is not a character record. Midjourney states that seeds cannot save a specific character or style, have relatively little influence on the finished image, and may behave differently across sessions. Its seed documentation directs consistency work toward reference features instead.

Keep the identity block unchanged and place scene instructions below it:

[Paste the unchanged CHARACTER LOCK block.]

SCENE
Spread ID:
Action:
Setting and time:
Camera distance and angle:
Characters visible:
Required props and continuity state:
Expression and pose:
Clear low-detail area reserved for text:
Final aspect ratio:

Save the exact scene prompt and reference files beside each spread ID. If one sleeve, shoe, or prop is wrong, use a focused edit when the tool supports it. OpenAI warns that a selected edit may extend beyond the highlighted region, so compare the full revised image with the approved version before accepting the repair.

Keep every story word outside the generated pixels

An image model may draw readable words, but the result is raster content. It cannot reflow when an editor changes a sentence, and correcting one word may require a new image edit. Generate the art without narration, page numbers, title lettering, or copyright-page copy. Add those elements in InDesign, Affinity Publisher, Scribus, Canva, or another application that keeps type editable.

Maintain one approved manuscript source. Give each text block the same spread ID used in the pagination map, then paste from that source into the layout. Paragraph styles should control font, size, leading, alignment, and spacing across the book.

For a sign, label, or shop window that needs exact wording, generate a blank surface with the required shape and perspective. Place the words separately in the layout. The spelling remains editable, and the type can be checked for contrast and safe-area placement without redrawing the scene.

Use fonts licensed for the intended publication and confirm that they can be embedded. KDP requires embedded interior fonts. Adobe's print-PDF guidance also calls for checking missing fonts, low-resolution graphics, modified links, bleed, color settings, and overset text before export.

Compose each image for its final text box and crop

"Leave room for text" does not define a layout. Specify the side, approximate proportion, and visual character of the empty area. One useful instruction is: "Reserve the upper-left quarter as pale sky with no faces, branches, or high-contrast objects." Check the generated scene with the actual text frame, because an area that holds two lines may fail when the approved copy needs six.

Place every image at its final crop and scale before approving it. Check:

• the character against the frozen reference sheet;
• the illustration against the manuscript line and continuity state;
• the text frame against the safe area and image contrast;
• the full-bleed background beyond each required trim edge;
• effective resolution after the final crop; and
• the complete spread with the center gutter visible.

Revise an image that cannot satisfy both the story composition and the printer's geometry. Stretching or repeatedly trimming the layout can alter proportions or remove a clue the manuscript expects the reader to see.

Review the complete sequence on one contact sheet

A contact sheet exposes changes that look harmless on isolated pages. At thumbnail size, check palette shifts, repeated camera angles, uneven text density, and whether each page turn reveals the intended action. At full size, compare recurring details with the reference sheet.

Inspect each spread in the same order:

1. Compare face shape, eye color, hair or fur pattern, body proportions, and apparent age.
2. Check outfit construction, one-sided marks, accessories, and shoes.
3. Confirm the number, color, ownership, and location of story props.
4. Carry forward weather, time, dirt, damage, or another state acquired earlier in the story.
5. Check movement direction across the page turn.
6. Read the exact manuscript line against the pictured action.

Record corrections by spread ID. "Restore the two yellow buttons from the approved reference" gives the editor an observable result. "Make the character consistent" does not identify the failed detail or tell the reviewer when the repair is complete.

Build the cover after the interior page count is final

The cover is a separate print file whose width depends on the bound book. On KDP, spine width changes with page count, ink, and paper. The KDP cover calculator generates a PDF or PNG template from the selected binding, color, paper, trim, reading direction, and page count.

Finish the interior PDF first, then generate the cover template and keep it as a locked guide layer. KDP's paperback cover instructions require one PDF containing the back cover, spine, and front cover. The current guidance says printed spine text needs more than 79 pages, while Cover Creator lists a minimum of 80. Leave spine text off a shorter KDP picture book.

Preflight the PDF and approve a physical proof

Export the interior in reading order as single pages. For KDP, omit crop and trim marks, embed the fonts, and flatten transparent objects and layers in the native file. Another printer may ask for a different PDF standard, bleed treatment, or color profile, so apply the chosen provider's preset.

Run a preflight before upload. Confirm page dimensions, page count, bleed, margins, image resolution, embedded fonts, updated links, and the absence of overset text. Adobe's packaging guidance recommends resolving preflight errors before collecting the native document, linked graphics, fonts, report, and print PDF.

Review the uploaded book in the platform's previewer, then order a physical proof. KDP's upload and preview instructions state that Print Previewer checks the paperback files for errors and that authors can order a printed proof. Inspect the proof under ordinary reading light. Check small type, gutter loss, trim variation, dark shadow detail, color shifts, and any spread that becomes hard to read when the book is held naturally.

If the book goes through KDP, disclose AI-generated text, images, or translations during setup. KDP's content guidelines count AI-created cover and interior artwork as AI-generated even after substantial edits. They also leave compliance with copyright, trademark, privacy, publicity, and other applicable rights with the publisher. Keep font, stock-asset, source-image, and model-use licenses in the project record.

Release the files only when the character sheet, locked manuscript, contact sheet, interior PDF, cover template, correction log, platform preview, and printed proof all describe the same version of the book.