Character Turnaround Sheets: Keep the Design, Change Only the View

A character turnaround sheet is one original design shown from the front, side, and back on a clean field. It is not a squad poster, not a copyrighted hero, and not a stat card the model is allowed to invent. Lock face, hair, silhouette, outfit, palette, and props first. Change only the camera. Leave nameplates, lore, and numbers empty if those words must be exact. Google’s Nano Banana 2.1 model card, updated 6 October 2026, documents up to 14 reference images and a fidelity band of about four characters and ten objects. That is a ceiling, not a brief to fill the page. ChatGPT Images 2.5, launched 8 September 2026, is a different family that keeps a subject more stably while you change one element. FreeGPTImg is a GPT Image 2 workspace: 20 daily free credits, no credit card required. Overlay real labels last.

Draft the sheet in FreeGPTImg. Run a structured front/side/back pass in the three-view character design sheet tool, explore the same lock in the AI character design sheet template, stay on GPT Image 2 for the loop in front of you, invent the first neutral front in AI text-to-image, change only the camera from a locked hero in image to image, try a stylized pass in the AI anime generator, and open Nano Banana Pro only when that is the label on the screen. Do not paste a screenshot of someone else’s character sheet.

What a turnaround actually has to do

Concept artists, indie game teams, and illustrators use a turnaround so later frames still show the same person. Front answers “who is this.” Side answers “how thick is the hair, the coat, the bag.” Back answers “does the design survive without the face.” If the side view grows a new haircut, you do not have a sheet. You have three cousins.

Google’s image-generation docs now treat Nano Banana 2.1 (gemini-nano-banana-2.1) as the high-efficiency default for new Gemini image jobs, with stronger multi-turn consistency than Nano Banana 2. DeepMind’s Flash page lists subject consistency and mask-based editing among the 2.1 gains. All Gemini outputs carry a SynthID watermark. That helps a keep-subject loop. It does not license a crowd, a copyrighted mascot, or a nameplate the model spelled.

ChatGPT Images 2.5 is not this family. Use it the same way you would use GPT Image 2: one original character, one crop, one noun. FreeGPTImg currently exposes GPT Image 2, Nano Banana 2, and Nano Banana Pro generator pages. Write the name on the screen into the file. Do not stamp a GPT Image 2 sheet as 2.1.

A four-character band is not a party scene. If you need a protagonist and a small companion, that is already two identities to lock. A “guild,” “squad,” or “battle lineup” is how you leave the band and how hair color starts drifting.

Name the sheet before you generate

Weak files fail because they ask for a turnaround, an expression grid, a weapon list, a logo, and a cinematic keyframe in one pass. Pick one job.

Job	What you export	Keep	Change one thing	Wrong move
Neutral front lock	One standing front, even light, empty labels	Face, hair, body, outfit, palette, props	Nothing yet	Action pose, extra people, invented stats
Three-view turnaround	Front, side, back of the same original character	The front lock, same costume construction	Camera / view only	New haircut on the side, missing bag on the back
Expression grid	Same head, empty caption bands	Bone structure, hair, eye color	Mouth and eyelids only	New hairstyle per cell, readable fake names
Outfit variant	Same body, one costume change	Face and proportions	Wardrobe only	A different person wearing “the same vibe”
Pose sheet	Same design, readable silhouettes	Costume seams and props	Pose only	Weapons you did not specify, extra characters
Realistic concept lock	Photoreal study of an original adult design	Same identity rules as the stylized lock	Medium or lighting	Celebrity likeness, copyrighted armor, children

Keep nameplates empty. Models still invent letters. OpenAI’s image-generation guidance still treats precise in-image text and packed layouts as weak spots. Google says Gemini works best if you generate the words first, then ask for an image that contains those words. For a sheet, that means: no stats, no “Lv. 12,” no lore paragraph in the pixels. Overlay in an editor after the views match.

Do not prompt copyrighted characters, trademarked hero suits, or a celebrity “as” your ranger. Photorealistic AI people may need a synthetic-performer tag in IPTC/XMP on some platforms; apply that in an editor. Prefer an original adult or a clearly stylized non-human. Never generate a child for a turnaround you plan to reuse in ads.

Clean generation workflow

Use the same sequence in the three-view tool, a GPT Image 2 tab, ChatGPT Images 2.5, or a Gemini 2.1 session. The rule is the lock, not the logo on the chat.

1. Write five lines. Original identity (face, hair, eye color, build), costume construction, one prop, sheet job (front lock / three-view / expression), model name on the screen.
2. Generate a neutral front first. Even light, plain background, standing, hands visible, empty labels. This file is the hero. Do not start with a three-quarter action shot.
3. Save a character lock paragraph. Short enough to paste. Face, hair length, palette, seams, prop placement. If a stranger cannot redraw from it, it is too vague.
4. Write keep / change / do-not. Keep the lock. Change only the camera, or only the mouth, or only the coat. Do not add a second character, a logo you do not own, or stats.
5. Build the turnaround from the front file. Image-to-image: “Keep this exact person. Change only to a side view.” Then a separate pass for the back. One noun per turn.
6. Generate three options on the same crop. Switching from a single front to a full three-view board between option A and option B is a new job.
7. Overlay real words last. Character name, game title, polycount, and lore never belong in the generator if they must be exact.
8. Inspect at 100%. Check extra fingers, mirrored scars that jump sides, missing props on the back, melted badges, and whether SynthID or Content Credentials are acceptable for that pipeline.

If you are on Gemini, 2.1 can take multiple reference stills in one request. Use them as the same person from useful angles, not as a moodboard of other people’s art. Google’s documented band is up to 14 references, with about four characters and ten objects held consistent. Stay inside one identity unless you have a real second character to lock.

Prompt formula

Original character turnaround, not a copyrighted hero and not a squad.
Model: [GPT Image 2 / ChatGPT Images 2.5 / Nano Banana 2.1 / Nano Banana Pro / name on screen].
Identity lock: [face shape, hair length and color, eye color, build, age band as adult or stylized non-human].
Costume lock: [silhouette, seams, palette, one named prop and where it sits].
Job: [neutral front lock / front-side-back turnaround / expression grid / outfit variant].
Crop: even full-body, plain [paper gray / studio white] background, character large in frame.
Change only: [camera to side / camera to back / mouth and eyelids / wardrobe].
Text in image: none. Empty nameplate and empty stat band if a board is required.
Do not add: extra characters, copyrighted costumes, celebrity likeness, children, logos I do not own, weapons not in the lock, readable stats.

Prompt examples

Neutral front lock (save this before any turnaround)

Original adult park ranger, not a celebrity and not a copyrighted mascot.
Oval face, short wavy black hair, brown eyes, medium brown skin, athletic build, calm closed-mouth expression.
Costume: olive field jacket, cream shirt, dark brown trousers, leather belt, scuffed hiking boots, small brass compass on a neck cord, no armor.
Standing front view, even studio light, plain paper-gray background, hands visible, empty labels.
Do not add a second person, a logo, or text.

Three-view from the locked front (image-to-image)

Image-to-image: keep this exact ranger, face, hair, jacket, compass, and boot color.
Build a clean three-view character reference sheet: front, side, back, same height, same costume construction.
Plain paper-gray background. Empty nameplate. No extra people. No readable stats.
Change only the camera. Do not restyle the design.

Side view only, one noun

Keep this exact ranger. Change only to a full-body left-side view.
Hair length, jacket hem, compass, and boot tread must match the front.
Plain background, empty labels, no second character.

Expression grid, empty captions

Same ranger head and shoulders, identical hair and jacket collar.
Six expressions in a clean grid: calm, surprised, tired, focused, amused, concerned.
Empty caption bands under each cell. No names, no numbers.
Do not change the haircut or eye color.

Review the pixels before you reuse the lock

Zoom to 100%. A thumbnail hides a missing compass on the back view.

• Identity honesty. Same face, hair length, eye color, build, and age. If a stranger would think the side view is a sibling, discard it.
• Costume honesty. Seams, pockets, boot height, and the named prop survive every view. A bag that vanishes on the back is a failed sheet.
• Empty type. Names, stats, titles, and lore stay empty of anything you need people to read.
• Rights honesty. Original design only. No copyrighted armor, no celebrity jawline, no children.
• Provenance. Gemini stills may carry SynthID. Downstream tools may keep, strip, or flag credentials. Match the file to the pipeline.

Limitations you should not skip

FreeGPTImg is a GPT Image 2 workspace with 20 daily free credits and no credit card required, plus live Nano Banana 2 and Nano Banana Pro pages. That is not a claim that FreeGPTImg runs gemini-nano-banana-2.1, that a three-view will be production-ready without a check, or that you own commercial rights in every territory. Match the label on the page you opened.

Do not promise that 14 references will hold a crowd. Google documents a four-character / ten-object band on 2.1. Treat it as a limit. Do not prompt a copyrighted roster “for practice.” Do not put live stats into the generator. Google is retiring Nano Banana 2 (gemini-3.1-flash-image) on 29 October 2026 on the Gemini API; new Gemini jobs should use 2.1, but do not relabel a GPT Image 2 file to match that calendar.

Models still invent extra people when you ask for a “party,” a “lineup,” or “more interesting.” Shorten the brief.

FAQ

Can I generate a usable front, side, and back in one image?

Sometimes as a first board, rarely as a finished lock. Generate a neutral front, inspect it, then change only the camera. If the combined board is the deliverable, still start from a locked front so the side and back have a source of truth.

What does Google’s four-character / 14-reference band mean?

On Nano Banana 2.1, Google documents up to 14 reference images, with about four characters and ten objects held consistent. That is a ceiling. A turnaround of one original person should use one identity. Extra references should be the same person, not a pile of unrelated art.

Should the character’s name and stats be in the picture?

No, not if they must be exact. Leave nameplates and number bands empty. Overlay the real name, title, and numbers in an editor after the views match.

Does FreeGPTImg run Nano Banana 2.1?

Use the label on the page you opened. FreeGPTImg currently exposes GPT Image 2, Nano Banana 2, and Nano Banana Pro. Write that name into the file. Do not invent a 2.1 badge because a Gemini doc used that id.

Can I make a turnaround of a famous game character?

Not as a sheet you reuse. Do not prompt copyrighted costumes, trademarked silhouettes, or celebrity likeness. Build an original lock. If a platform requires a synthetic-performer tag for photoreal people, apply it in an editor.

Why did the side view change the haircut?

The prompt asked for a “cool new angle,” a “full character sheet,” or “more detail” without locking hair length and the front file. Shorten it. Keep this person. Change only the camera. Generate again from the locked front.

Sources

• Google AI for Developers: Nano Banana image generation
• Google DeepMind: Nano Banana 2.1
• Google DeepMind: Nano Banana 2.1 model card
• Google DeepMind: Model cards
• Google DeepMind: SynthID
• OpenAI: Introducing ChatGPT Images 2.5
• OpenAI: Image generation
• C2PA: Content Credentials

Make one three-view lock on FreeGPTImg

Write five lines: the original identity, the costume, the one prop, the sheet job, and the model name on the screen. Open FreeGPTImg, run the board in the three-view character design sheet tool or the AI character design sheet template, stay on GPT Image 2 when that is the loop in front of you, invent the first front in AI text-to-image, change only the camera in image to image, try a stylized pass in the AI anime generator, and use Nano Banana Pro only when that label is on the page. Download the image. Leave the nameplate empty until the three views match.