AI Character Sheet: Keep One Character Consistent Across Views

Published on October 15, 2026 by Steven Jones

An AI character sheet shows one character from the front, side and back so you can reuse their appearance in different illustrations. The challenge is not producing three attractive figures. It is keeping the same face, haircut, clothing and distinguishing marks after the character turns, moves or enters a new setting.

We built one turnaround in the DIY AI image generator and reused it as the reference for three separate GPT Image 2 scenes. Below are the actual outputs, our prompts, the controls used, and a checklist for catching identity drift before it spreads across a storyboard.

Test verdict: The character remained recognisable in a reading room, a running profile and a rear-view walking scene. Her red bob, teal streak and mustard jacket survived. Smaller costume details drifted, particularly the cargo pockets. This was useful for reference consistency, not a guaranteed identity lock.

What our four-image character test proves

We generated the reference sheet once, then branched all three scenes directly from that original file. We did not feed the reading-room result into the running scene, or the running scene into the garden scene. This avoids passing a mistake from one generation into the next. The test demonstrates three reference-led outputs, not a claim that every future frame would remain identical.

Image	What held up	What needs checking
01 – Turnaround	Matching outfit palette, build and footwear across three views	Sleeve-patch placement is not equally clear from every angle
02 – Reading room	Face shape, copper-red hair, teal streak and jacket	Lower legs and boots are outside the frame
03 – Running	Side silhouette, teal streak, jacket, trousers and burgundy boots	Dynamic pose changes perceived limb proportions
04 – Garden	Haircut, jacket silhouette, colours and boots from behind	Extra cargo-pocket detail appears compared with the master

The important failure is visible in panel 05: the garden image introduces additional trouser-pocket detailing. A casual viewer would still recognise Mara, but someone drawing the next scene could accidentally treat the new pocket arrangement as canonical. This is why a usable character sheet needs an approval rule, not just a similarity impression.

---

---

Step 1: Define identity separately from the scene

Create a short character specification before asking for poses. Prioritise features that are recognisable at different distances: head shape, hairstyle, build, clothing silhouette, and a few distinctive marks. Give left and right from the character’s perspective. A sleeve emblem on the wrong arm counts as a continuity error, even if it is the correct colour.

Our fictional character, Mara Vale, has warm olive skin, an angular nose, a chin-length copper-red bob with a teal streak, a cropped mustard jacket with an indigo collar and circular orange sleeve patch, dark teal cargo trousers, a tan belt and burgundy boots. Those characteristics stay fixed. Camera angle, weather, pose, background and lighting are scene variables.

Avoid making every tiny feature mandatory. The eyebrow nick and earring are helpful at close range but cannot be judged in a back view or distant action shot. Treat details hidden by the camera as unobservable, not successful matches.

Step 2: Generate front, side and back views together

Use a neutral turnaround as your first image. Full-body figures, level camera height and even lighting reveal discrepancies that dramatic shadows would conceal. A true side profile is more useful than a flattering three-quarter pose when you later need the character running across the frame.

Create one 2D illustrated character turnaround showing exactly the SAME adult character in three FULL-BODY views: straight front, strict left-side profile and straight back. Match height, facial anatomy, hair shape, outfit cut, colour and footwear. Character: Mara Vale, warm olive skin, angular nose, copper-red chin-length bob with one teal streak, mustard cropped jacket with indigo collar, circular orange sleeve patch, dark teal cargo trousers, tan belt and burgundy boots. Neutral standing poses, level camera, matching scale, evenly lit plain off-white background. One person repeated in three reference views; no scenery, captions or extra figures.

Inspect all three figures before saving the master. Compare head-to-shoulder ratio, boot height, jacket hem, sleeve markings and which trouser leg carries the cargo pocket. In our original sheet, the basic costume reads consistently, but the patch is easier to see from the side and rear. Where an attribute is ambiguous, clarify it in the written identity brief instead of assuming the image settled it.

Step 3: Reuse the approved reference without accumulating errors

For our three scene tests, we supplied the entire original turnaround as a single reference image each time. That is an important methodological detail: we did not pre-crop separate angles for the published test. A multi-panel image can sometimes cause a generator to repeat the character or copy the sheet layout, so explicitly request one figure in one scene. None of our three finished outputs repeated the turnaround layout.

For longer production, export approved front, side, and back crops as separate JPGs, then compare them with the full-sheet approach on the same prompts. Creator discussions repeatedly report mixed outcomes when comparing a multi-view sheet with one close reference. The useful question is which input keeps your model more faithful at the camera angles you actually need, rather than assuming more references always improve the result. Our three-scene test did not establish a winner between those two approaches.

Keep the original as mara-master-v1.jpg and use a predictable revision name for corrected views. Do not promote a visually impressive but inaccurate scene into the next generation’s identity source.

Step 4: Challenge the character with three different shots

Changing only the backdrop is too easy a test. We deliberately changed camera angle, lighting and body movement. Add the same instruction to every scene prompt: “Use the attached turnaround for character identity and clothing only. Preserve the same character. Show one figure, not a reference-sheet collage.”

Front three-quarter: reading room

Use the attached Mara Vale character turnaround as the exact identity and costume reference. Make one 2D graphic-novel scene: Mara leaning over a map in an old railway-station reading room, medium three-quarter front view, warm light through a window. Preserve her copper-red bob, right-side teal streak, angular face, mustard cropped jacket, indigo collar, sleeve patch and dark teal trousers. One Mara only. No triptych or text.

Observed: The face and hairstyle remain recognisable, including the teal streak. This framing crops out the boots, so it is not evidence that footwear remained correct in the reading-room result. See image 02 and the facial comparison in image 05.

Strict profile: running through rain

Use the same original Mara Vale turnaround. Draw one full-body LEFT-SIDE profile of Mara running across a rain-slick futuristic street. Keep her facial silhouette, red bob with teal streak, mustard cropped jacket, indigo collar, orange sleeve patch, teal cargo trousers and burgundy lace-up boots. Cinematic night reflections, dynamic motion. Do not change her identity or produce multiple character panels.

Observed: The jacket, hair and boots remain readily identifiable from the side. The extended running stride makes it harder to compare body proportions fairly with the neutral standing reference. Inspect the nose profile and clothing cut rather than judging proportions from the overall silhouette alone. See image 03.

Rear three-quarter: walking away

Use the original Mara Vale turnaround as the fixed reference. Illustrate one scene of Mara walking away along a sunny elevated garden walkway, seen from a rear three-quarter angle while she looks back over one shoulder. Preserve her haircut, jacket back and collar, orange sleeve patch, dark teal trousers and burgundy ankle boots. One person, one landscape scene, no labels or panels.

Observed: The red bob, jacket length and burgundy boots survive the camera change. The trousers gain extra visible cargo-pocket detailing, which breaks the original wardrobe specification. That is a real continuity issue even though the picture works artistically. See images 04 and 05.

Step 5: Run an identity-drift check before keeping a shot

• Face: Nose profile, jaw, eye spacing and apparent age. Reject a new face rather than trying to describe it as the same character.
• Hair: Length, parting, distinctive streak and outline. Check only what that angle exposes.
• Clothing: Jacket hem, collar, sleeve patch, belt, cargo-pocket count and boot design. Small positional mistakes often survive a quick glance.
• Shot integrity: One character, correct camera angle, plausible limbs and no repeated turnaround panels.
• Rendering: Mark changes in line weight or painterly texture separately from changes in character identity.

Use three outcomes for each visible feature: match, changed or cannot judge. Do not pretend every characteristic can be scored in every shot. In this test, hairstyle and main costume colours match across scenes; fine wardrobe geometry needs correction, and close facial details, such as the eyebrow nick, aren’t verifiable in all three.

Your search, your sources

Make DIY AI a preferred source

See more of our reporting in Google Top Stories, AI Overviews and AI Mode.

Add DIY AI on Google

Which Studio image controls were actually available?

The master used DIY AI Studio’s GPT Image 2 text-to-image route in 16:9. Each subsequent scene used its GPT Image 2 reference-edit route with the saved master image, and a 3:2 output. This route rejected a 16:9 reference-edit attempt; the supported reference-edit shapes in our test were 1:1, 3:2, and 2:3. Do not confuse settings available for text-to-image with those available when attaching a reference.

Each completed image cost 4 Studio credits during our test, so the master plus three scenes consumed 16 credits. This is a record of one run, not a permanent price or an assurance that every rerun will succeed the first time. We did not use a seed lock, a numeric identity-strength slider, a trained character model, or face-lock control. OpenAI’s image prompting guidance recommends stating what an edit may change and what it must preserve, then inspecting the output.

A Studio workflow limitation also surfaced: the three completed reference-edit assets had working authenticated result links but did not appear in the usual Studio image asset list. We retrieved the files separately for review. Until that behaviour is resolved, confirm you can actually save your edited images before relying on the gallery as the project archive.

When is the sheet ready to use in a storyboard?

Approve the turnaround first, then accept only scene images that preserve the character’s fixed identity. If a scene adds pockets, switches a patch or changes the nose, correct that detail or regenerate from the master. Save the camera angle and prompt alongside each accepted scene so the next shot can reuse the same specification without inheriting an accidental redesign.

Once the identity references are reliable, use the AI storyboard generator guide for shot sequencing, framing and continuity between scenes. If you instead need a single forward-facing talking presenter, start with AI avatar prompts; a portrait for lip-sync has different constraints from a front/side/back turnaround.

Practical stopping rule: A recognisable face in three scenes is promising. A production-ready character reference also needs fixed, inspectable costume details. Treat the first repeated error as a reason to repair the reference workflow before commissioning a longer series.