To keep an AI character consistent, fix its look once and reuse it: make a character reference sheet (front, three-quarter and action views on a plain background), attach it to every illustration the character appears in, paste the same style description into every prompt, and redraw the pages that drift. A text description alone isn’t enough, because each picture is drawn from scratch.
Why AI characters drift
Unless it is given the earlier picture, an image model doesn’t know what the fox on page 3 looked like when it draws page 4. Given only words, it draws a new fox that fits the description, and “a small red fox in a yellow scarf” fits thousands of foxes. Readers notice quickly, and in a picture book the main character is on almost every spread and on the cover. What changes, and what helps:
| What drifts | Why | What helps |
|---|---|---|
| Face and proportions | Each picture is generated from scratch; a description fits many faces | A reference image of the character with every picture |
| Outfit details | Patterns, buttons and logos are the first details to change | A simple signature outfit, shown on the reference sheet |
| Colors | Scene lighting and style words shift the palette | A sheet drawn in flat, even light; colors named in the description |
| Art style | Each prompt is worded a little differently | One style paragraph, pasted word for word |
| Two characters in one scene | Traits blend when references and descriptions overlap | One sheet per character, and only the sheets of the characters in that scene |
| Relative size | Nothing tells the model how big each character is | Sizes stated in the description |
Design a character you can repeat
Some designs survive dozens of illustrations better than others. A clear silhouette, two or three main colors and one signature item (a scarf, a hat, a backpack) are easy to repeat. Tiny patterns, text or logos on clothes, jewelry and complicated hair are hard, and they are the details children point at when they change.
Write the look down once, in a fixed order, and reuse it unchanged:
Pip: a young red fox, small and round.
Face: short snout, big amber eyes, white chest patch, black-tipped ears.
Fur: bright orange-red, white tip on the tail.
Outfit: a plain mustard-yellow knitted scarf, always worn.
Size: comes up to the shoulder of Bramble the badger.Species or age, build, face, hair or fur, colors, outfit and distinctive marks: the same fields every time. If a detail doesn’t matter to the story, leave it out rather than let it change from page to page.
Make a character reference sheet
A reference sheet is the image every later picture is matched against, so it has to be clean and unambiguous:
- Three views side by side: front view (full body, neutral pose), three-quarter view (full body) and one action pose that shows the character’s personality.
- Identical in every view: face, proportions, colors, clothing and marks, drawn at the same scale, with nothing cropped.
- Plain, light background and even lighting, so the true colors are clear. No props, scenery or other characters.
- No text: no name, labels, arrows or color swatches, which can end up in the pages.
- In the book’s final style and color mode. A sheet in watercolor won’t anchor a book in flat vector art, and a coloring book needs a black-and-white line-art sheet.
Make one sheet per main character, check it carefully before any page is drawn, and keep it: it is also what brings the same character back in book two of a series.
Lock the style
Style drift is as visible as character drift. Write one style paragraph and paste it, unchanged, into every prompt, including the cover. Keep the same tool and model for the whole book; different models draw the same words differently.
Soft watercolor picture-book illustration, warm muted palette, visible paper texture,
loose ink outlines, gentle daylight, calm background areas. No text or letters.A style reference image helps if your tool accepts one, but restate the style in words as well: references alone can drift. Browse the 40 illustration styles for wording that suits children’s books.
Prompt tips
- Say what each attached image is. “Image 1 is the character sheet of Pip: match the face, proportions, colors and scarf exactly, but ignore its poses, layout and background.” Without that, models can copy the sheet’s pose or white background.
- Attach only the characters in the scene. A sheet for a character who isn’t on the page can add them to the scene.
- Describe the scene, not the look. The look comes from the sheet and the fixed description; the prompt adds the action, the place, the emotion and the camera.
- Keep text out of the art. Set the words as real text in the layout, so they stay sharp and editable, and ask for a calm area where the text will sit.
- Mind the edges. For full-bleed pages, keep faces and key objects well inside the page: we keep them at least 8% from every edge and 12% from the binding side, so nothing is lost in the trim or the gutter.
- Keep a good earlier page at hand. Attaching a page you’re happy with, as a style reference, helps later pages match its colors and light.
A page prompt put together this way:
Image 1 is the character sheet of Pip: match face, proportions, colors and scarf exactly;
ignore its poses and background.
Scene: Pip kneels at the edge of a stream at dawn, reaching for a paper boat, surprised.
Wide shot, Pip in the lower left third, calm sky in the upper half for the text.
[style paragraph, word for word]
Avoid: text, letters, extra characters, a second scarf.DIY methods and their limits
Every method below can work for a short book. None removes the need to look at every page.
| Method | Good for | Limits |
|---|---|---|
| Fixed written description in every prompt | Species, main colors, outfit | Faces and small details still change from picture to picture |
| A reference image per character in an image tool that accepts image inputs | Face, proportions, colors | Poses and fine details can still drift; scenes with several characters are the hardest |
| Midjourney Omni Reference (V7) | Putting one character or object from an image into new scenes | One reference image only, 2× the GPU time of a regular image, and Midjourney says intricate details like freckles or logos may not match |
| Editing a detail on one page (inpainting, regional redraw) | Fixing a scarf color or an eye without redrawing the page | Manual and page by page |
| Training a small custom model on your character | A strong identity lock over many images | Technical setup, and you need a consistent set of images to train on first |
| Drawing the art yourself or hiring an illustrator | Full control, and no AI images to declare | Time or budget |
About Midjourney (checked September 28, 2026): Omni Reference works with version 7, takes a weight from 1 to 1,000 (default 100), and for several characters its docs suggest one image that contains them all. For version 8, Midjourney points to its Edit Model, which takes up to 4 reference images. Whatever the tool, you still assemble the pages, the text, the print PDF and the cover yourself.
How kdpbook.io does it
kdpbook applies the method above automatically, inside a tool that also builds the KDP files:
- Cast. Each main character gets a name, a role and a written look, or up to 3 photos or drawings that are described and redrawn in the book’s style.
- Sheets first. A three-view reference sheet is drawn for up to 4 main characters and locked before any page is painted. A sheet costs 15 credits, a little more when it is drawn from your pictures. You can pause after the plan to review the story and the page plan first.
- Every page gets its references. Each illustration is painted with the sheets of the characters in that scene, up to 2 earlier pages for continuity, and the style restated in words, in one of 40 styles or your own (described in words or from up to 3 style images).
- Fix what drifts. In the editor, redo any illustration with the same references, or without them for a fresh start. Redraws use credits.
- Series. Character sheets from your finished books stay in your library, so a character can come back in the next book.
The limits are real: image models don’t follow references perfectly, so a small detail can still change on a page, and if a sheet can’t be drawn the pages are painted from the written description (you aren’t charged for the failed sheet). You still read and check every page. New accounts get 2,000 free credits once, with no card required. Free-plan books carry a kdpbook.io preview watermark and are shared in the community; any paid plan exports clean files. A 32-page picture book is estimated at about 1,624 credits, within the free credits. For other tools, see AI children’s book generators compared.
Print specs for a picture book
Consistent art still has to fit KDP’s print rules. The common setup, and the default of the AI picture book generator, is an 8.5 × 8.5 in premium color paperback of 32 pages:
- Bleed: full-page art runs past the trim, so each interior page is 8.625 × 8.75 in.
- Resolution: KDP asks for 300 DPI at print size, so a full-bleed page needs about 2,588 × 2,625 pixels. Many image tools output less, so check the size and upscale before layout.
- Pages: at least 24. At 32 pages the spine stays blank: KDP allows spine text only from 79 pages.
- Cost: on Amazon.com this book costs $4.20 to print (a flat fee up to 40 pages), so the minimum list price is $8.40. See how to price a book on KDP.
- Kindle: picture books are usually published as fixed-layout eBooks, which keep every page’s design; see fixed-layout Kindle picture books.
The step-by-step version is in how to publish a children’s picture book on KDP, and word counts by age are in children’s book word count.
AI disclosure and rights
KDP defines AI-generated content as text, images or translations created by an AI tool, even if you edited them substantially afterwards, and asks you to declare it when you publish or republish. Images include the cover and the interior art. If you designed the character yourself but an AI painted the pages, the pages are AI-generated images. Content you created and only refined with AI tools is AI-assisted and doesn’t need to be declared.
For a book whose illustrations were all generated, the images answer is usually “Many AI-generated images, with minimal or no editing”. KDP also makes you responsible for intellectual property: don’t recreate trademarked or well-known characters, and base a character on a real person only with their agreement. Details in the AI disclosure guide and can you publish AI-generated books on KDP?