How to Write an AI Character Reference Sheet Prompt from Your Story Bible

Why should your story bible include character images?

A story bible exists to stop continuity errors — the heroine whose eyes drift from green to hazel between books, the scar that switches forearms. A reference sheet does the same job for everything visual, and it gets used more often than you'd expect:

• Cover art and marketing. A cover artist — human or AI — needs a visual spec, not three paragraphs of scattered description.
• Artist commissions. Character artists routinely ask for reference images. A turnaround sheet answers costume, palette, and silhouette questions in one attachment instead of a dozen back-and-forth emails.
• Series continuity. When book three's promo art is made two years after book one's, the sheet is the ground truth you check against.
• Your own drafting. Writers describe characters more consistently once they've seen them. It's much easier to notice that your "weathered captain" reads as thirty-five in your head but fifty on the page.

This works with almost no extra effort because story-bible appearance sections and image prompts are the same data in different formats. You're not writing something new — you're reformatting fields you filled in months ago.

How do story-bible appearance fields map to prompt fields?

Almost one-to-one. Here's the mapping from a typical character-bible template to prompt text:

• Race/species → race. Lead with it; it sets anatomy for everything after.
• Age and build → body descriptors. Be numeric-adjacent: "in her fifties, broad-shouldered, heavy build" beats "older, imposing."
• Hair, eyes, skin → the identity core. These three plus face shape are what image models latch onto for likeness. Put them early — models weight early tokens more heavily and quietly drop details buried at the end of long prompts.
• Distinguishing marks → immediately after the identity core. Scars, tattoos, heterochromia, a missing finger. These are your continuity anchors, so never let them sink to the end of the prompt.
• Occupation → character type. "Apothecary" or "mercenary" pulls in an entire costume-and-context vocabulary the model already knows.
• Wealth and class → materials. This is the mapping writers miss. "Modest tradesperson" doesn't render; patched wool and coarse linen versus silk and gold thread absolutely do.
• Personality → grooming and resting expression. "Brave" does nothing visually. "Meticulously kept uniform, guarded expression" does.

Skip everything else. Backstory, motivations, fears — none of it renders, and it dilutes the tokens that do. The generator structures this same mapping as ~25 trait fields, which is useful when you want the wealth-to-materials translation done for you.

How do you prompt a front/side/back turnaround sheet?

Professional animation model sheets show a character in four to five views — front, three-quarter front, profile, three-quarter back, back — in a neutral pose under flat light. AI models reliably fit three or four views in one image; ask for more and views start merging or contradicting each other.

The structure that works: your appearance paragraph first, then a turnaround suffix that names each view explicitly and demands consistency.

Character reference sheet of Maren Voss, a weathered human sea captain in her fifties, silver-streaked black hair in a tight braid, gray eyes, a pale scar through her left eyebrow, broad shoulders and a heavy build. She wears a salt-stained navy wool greatcoat with brass buttons over a coarse linen shirt, cracked leather boots. Full body turnaround: front view, three-quarter view, side profile, and back view of the same character in a relaxed neutral standing pose. Flat, even lighting, plain light-gray background, consistent proportions and colors across all views, clean digital painting style. --ar 16:9

Four rules make the difference:

1. Name every view. "Front view, three-quarter view, side profile, back view" — vague phrasing like "multiple angles" produces four random poses of four slightly different people.
2. Say "the same character" explicitly. It measurably reduces trait drift between views.
3. Use a wide aspect ratio — --ar 16:9 or wider in Midjourney — so all full-body views fit side by side without cropping feet.
4. Ask for a relaxed neutral stance, not a T-pose. T-poses come from 3D rigging workflows, and image models trained mostly on natural photos often mangle the outstretched arms.

Expect to generate three or four candidates and keep the one where the views agree best. Perfect cross-view consistency in a single generation is still luck; you're picking the best liar, then promoting it to canon.

Why generate a neutral reference sheet before any scene art?

Because mood destroys information. Put your captain in candlelight and everything shifts amber — you can no longer tell whether her coat is navy or black, whether her skin is fair or ruddy. Generate the dramatic scene first and promote it to reference, and you've canonized lighting artifacts as character traits.

Animation studios solved this decades ago: the model sheet is drawn flat and approved before a single scene is animated, precisely so there's an unambiguous record of true colors and proportions. Your reference sheet plays the same role. Flat, even light — the closest studio setup to soft window light — shows the actual palette of skin, hair, and costume with nothing baked in.

There's a second, more practical reason: the neutral sheet is what you feed into image-reference features later. Midjourney's Omni Reference, Stable Diffusion's img2img and IP-Adapter, ChatGPT's image uploads — all of them transfer whatever is in the reference image, including its lighting. A reference shot in storm light drags cold blue shadows into your sunny tavern scene. A flat-lit reference transfers identity and costume cleanly and lets each scene prompt set its own mood.

Order of operations, then: neutral turnaround first, lock it as canon, and only afterward spend generations on the moody stuff.

How do you keep the character consistent in later scene images?

Two mechanisms, used together: a verbal identity block and an image reference.

The identity block is a 25–40 word excerpt of your sheet prompt — race, age, build, hair, eyes, marks, signature garment — that you reuse verbatim in every scene prompt. Verbatim matters: "silver-streaked black hair in a tight braid" and "graying dark hair, braided" are the same sentence to you and two different characters to the model. Keep the block in your story bible next to the appearance fields it came from.

Maren Voss, a weathered human sea captain in her fifties, silver-streaked black hair in a tight braid, gray eyes, a pale scar through her left eyebrow, wearing a salt-stained navy wool greatcoat with brass buttons. Three-quarter portrait at a rain-lashed ship's wheel, storm light with cold rim highlights and deep shadows, muted desaturated palette, digital painting style.

The image reference points the model at your neutral sheet. In Midjourney V7 and later, that's --oref with your sheet's URL; --ow (omni weight, 0–1000, default 100) controls adherence — raise it toward 400 when the face or costume drifts. The older --cref parameter only works in V6 and Niji 6; V7+ rejects it. In Stable Diffusion, use img2img or IP-Adapter with the sheet, or the CharTurner embedding plus ControlNet OpenPose when you need new turnarounds of the same character. In ChatGPT, attach the sheet image and instruct it to match the character exactly.

Identity block plus image reference beats either alone — the text pins the traits the reference transfer misses, and the reference pins the face that text can't fully specify. The full workflow, including seed reuse and drift correction, is in the consistent AI characters guide.

What does a real story-bible profile look like as a reference sheet prompt?

Take a typical character-bible entry, the kind a Reedsy or Novel Smithy template produces:

• Name: Ilya Renn
• Age: 34
• Race: half-elf
• Occupation: apothecary
• Build: slight, narrow shoulders
• Hair: copper-red, cropped short
• Eyes: amber
• Marks: ink-stained fingers; burn scar on right forearm
• Wardrobe: patched linen apron over a moss-green wool tunic; brass-framed spectacles
• Class/wealth: modest tradesperson

Every line maps. Half-elf and apothecary lead as race and character type. "Modest tradesperson" becomes the patched apron and the wool-not-velvet tunic. The burn scar and ink stains go right after the identity core because they're the continuity anchors. Only "34" gets softened — models handle "in his thirties" better than exact ages.

Character reference sheet of Ilya Renn, a slight half-elf apothecary in his thirties with narrow shoulders, cropped copper-red hair, amber eyes, and brass-framed round spectacles. Ink-stained fingers and an old burn scar on his right forearm. He wears a patched linen apron over a moss-green wool tunic, a cord of dried herbs at his belt. Full body turnaround: front view, three-quarter view, side profile, and back view of the same character in a relaxed neutral standing pose. Flat, even lighting, plain neutral background, consistent proportions and colors across all views, clean digital painting style. --ar 16:9

Notice what was left out: his motivation (find a cure for his sister), his fear of open water, his dry humor. None of it renders. If a personality trait must be visible, translate it — the dry humor could become "a faint skeptical half-smile," but on a turnaround sheet, keep even that minimal. The sheet's job is anatomy, palette, and costume; expressions belong in scene art, in a style you've locked here first.

Frequently asked questions

What views does a professional character turnaround sheet include?

Traditional animation model sheets show four to five views: front, three-quarter front, side profile, three-quarter back, and back, all in a neutral pose under flat lighting. For AI generation, three or four named views is the practical limit — front, three-quarter, side profile, and back. Asking for five or more in one image usually causes views to merge or show contradictory details.

What aspect ratio should I use for an AI turnaround sheet?

Wide. Use 16:9 as a default, or stretch to about 3:1 if you want four full-body views in a single clean row. Square formats force the model to stack or shrink views, which crops feet and compresses detail. In Midjourney, set it with the --ar parameter; in Stable Diffusion, set a wide resolution like 1536x640 before generating.

Does Midjourney's --cref parameter still work for character reference?

Only in older models. The --cref character reference parameter works with Midjourney V6 and Niji 6, but V7 and later reject or ignore it. Current versions use Omni Reference instead: add --oref with your reference image URL, and control adherence with --ow (omni weight), which ranges 0 to 1000 with a default of 100. Raise it toward 400 when facial features or costume details drift.

Can I put two characters on one AI reference sheet?

Avoid it. When two characters share one prompt, image models blend their traits — one character's hair color bleeds onto the other, scars migrate, costumes swap pieces. Generate a separate turnaround sheet per character, then compose group scenes later using each character's identity block and image reference. Sheets are cheap; untangling blended canon is not.

Is a T-pose or a natural standing pose better for AI character sheets?

A relaxed neutral standing pose. T-poses come from 3D rigging pipelines, and diffusion models trained largely on natural photography frequently distort the horizontally outstretched arms — extra joints, wrong lengths, merged hands. A neutral stance with arms at the sides renders far more reliably and still shows silhouette, costume, and proportions clearly from every view.

Should a character reference sheet include facial expressions?

Not on the turnaround itself — keep that neutral so proportions and palette stay readable. If you want expressions, generate a second sheet: a head-and-shoulders grid of the same character showing four to six labeled emotions, using the same identity block and your neutral sheet as an image reference. Animation studios separate turnarounds and expression sheets for the same reason.

Can I make a reference sheet from an existing image of my character?

Yes. If you already have one good portrait, use it as an image reference — Omni Reference in Midjourney V7+, img2img or IP-Adapter in Stable Diffusion, or an image upload in ChatGPT — and prompt for a full-body turnaround with named front, side, and back views. Keep your written appearance description in the prompt too; the text fills in whatever the single reference image doesn't show, like the back of the costume.

Can I send an AI-generated reference sheet to a human artist for a commission?

Usually, but say so upfront. Many character artists accept AI images as visual reference for likeness, costume, and palette, while some decline AI-adjacent work entirely — check their commission terms. As reference material, a turnaround sheet works well: it answers questions about silhouette, color, and costume construction in one image instead of a long description thread.