Home / Blog / Prompt & workflow guide

Prompt & workflow guide

Same character, every scene: keeping an AI character consistent across 4 models

Nov 13, 2026 · 8 min read · By the Pizyn team

The hardest prompt in AI image work is not a scene — it is a person you have already drawn once. A character sheet in words gets you most of the way inside one model, but the moment you switch models, or ask for a new scene, the face starts to wander. So we built one character, Mara, and ran her through the two methods that actually work: a written character sheet, and a reference photo fed to four different editors. Six generations, $0.19 in provider fees, first outputs kept — and every model held her face. Almost.

Mara, a 32-year-old wildlife photographer with curly auburn hair, round tortoiseshell glasses and a forest-green field jacket, standing in a pine forest clearing, generated with Seedream 5.5

How this test ran

Every frame below was generated on November 13, 2026 through this studio, first outputs kept, prompts quoted verbatim. The cast is one fictional character, Mara, a 32-year-old wildlife photographer. Six generations — two text-to-image and four reference edits — cost $0.19 in provider fees. The base portrait above is Seedream 5.5 at 1K, 4 credits; the four edits each took that portrait as their only reference and moved Mara to a new scene.

The test has two parts. First, the control: can the same words, with no reference photo at all, reproduce the same woman in a new scene? Second, the edits: when four different models are handed the same reference photo, which ones actually keep the face?

Step one: write the character sheet

Consistency starts before any reference image exists, because the first portrait has to come from text alone. The trick is to describe the character the way a casting director would: fixed, checkable traits, repeated word for word in every prompt. Not "a beautiful woman" — that is a different woman every time — but hair, glasses, freckles, jacket, camera. Our sheet, used verbatim in every prompt in this test:

"Mara, a 32-year-old wildlife photographer with curly auburn hair tied back in a low ponytail, round tortoiseshell glasses, freckles across her nose, wearing a forest-green canvas field jacket over a cream knit sweater, holding a vintage film camera with a leather strap around her neck"

Three rules make a sheet like this work. Pick traits a model can see: hair shape and colour, glasses, freckles, a specific jacket — not personality or backstory, which no image model can render. Keep it short enough to repeat: if the sheet is a paragraph, you will paraphrase it next time and the character will drift with your wording. Freeze the wording: copy-paste it, don't retype it. Our prompt guide makes the general point — concrete, repeatable nouns beat adjectives — and a character sheet is that point applied to a person.

The base portrait was the sheet plus a scene, Seedream 5.5 at 1K, 3:4:

"Editorial portrait photograph of Mara, a 32-year-old wildlife photographer with curly auburn hair tied back in a low ponytail, round tortoiseshell glasses, freckles across her nose, wearing a forest-green canvas field jacket over a cream knit sweater, holding a vintage film camera with a leather strap around her neck, standing in a pine forest clearing in soft morning light, photorealistic, shallow depth of field"

It landed first try: curly auburn hair in a low ponytail, round tortoiseshell glasses, freckles, the green jacket over cream knit, the film camera on its strap. That portrait is now the reference photo for everything that follows — save your best first portrait, because it becomes the anchor for the whole series.

The control: no reference, same words

Before spending anything on edits, we asked the obvious question: does the sheet alone reproduce Mara? Same model, same wording, new scene — Kyoto at night, no reference image:

"Editorial photograph of Mara, a 32-year-old wildlife photographer with curly auburn hair tied back in a low ponytail, round tortoiseshell glasses, freckles across her nose, wearing a forest-green canvas field jacket over a cream knit sweater, standing on a rainy neon-lit street in Kyoto at night, holding a vintage film camera with a leather strap, wet pavement reflections, photorealistic"

Mara standing on a rainy neon-lit street in Kyoto at night, generated from text alone with Seedream 5.5 and no reference photo — **The control · Seedream 5.5, text only, 1K · 4 credits.** No reference photo — the character sheet alone. She is recognisably the same woman, but the face is a close relative, not a twin: the pose flipped to a three-quarter view, the hair looser. Good enough for a series of illustrations; not good enough if a reader must believe it is the same person.

This is the honest result, and it cuts both ways. A detailed sheet does work inside one model: the hair, glasses, freckles, jacket and camera all survived the scene change, and a casual reader would accept this as Mara in Kyoto. But look at the face next to the base portrait and the drift is real — the angle changed, the features softened, the likeness is approximate. Text describes a type; a photo pins a person. That is the whole case for reference editing: the sheet gets you the character, the reference keeps the face.

Four edits, one reference photo

Each edit took the forest portrait as its only reference: keep this exact woman, change the scene. Four models, four scenes, first outputs kept.

Mara on a rainy neon-lit street in Kyoto at night, edited from the base portrait with Seedream 5.5 — **Seedream 5.5 edit · Kyoto rain · 1K · 5 credits.** Same street as the control — now with the reference photo attached. The face is the base portrait's face, the ponytail, glasses and jacket all intact, and the scene is richer: a rickshaw, wet pavement, neon that actually reflects.

Mara in a golden desert at sunset, edited from the base portrait with Nano Banana 2.1 — **Nano Banana 2.1 edit · Desert sunset · 1K · 8 credits.** The warm rim light is doing real work here — it relights her hair without redrawing her face. Freckles, glasses, jacket, camera strap: all carried over.

Mara seated on a wooden stool in a bright photography studio, edited from the base portrait with GPT Image 2.6 — **GPT Image 2.6 edit · Studio portrait · low · 8 credits.** The biggest scene change in the set — forest to seamless studio, standing to seated — and the face holds. It even kept the window-light shadows on the backdrop. Jeans appeared unprompted; a fair trade.

Mara standing on a misty mountain ridge at dawn, edited from the base portrait with FLUX 2 Pro — **FLUX 2 Pro edit · Mountain ridge · 2:3 · 10 credits.** The same woman, but the wardrobe wobbled: the jacket lost its brown collar and the camera strap crossed her chest instead of hanging at her neck. The face is right; the outfit is a near miss.

Put the Seedream edit next to the control and the value of the reference is visible in one glance: same scene, same model — the edit is Mara, the control is someone who looks like Mara.

Scores: who kept Mara

Judged on face, hair, glasses and outfit against the base portrait — first outputs, no cherry-picking:

Seedream 5.5 edit — 5/5. Near-identical: face, ponytail, jacket and camera all survive, and the new scene is fully rendered rather than pasted in. Also the cheapest edit in the test.
Nano Banana 2.1 edit — 5/5. Near-identical, and the best relighting: the sunset wraps her hair in warm light without changing her features. Its 14-reference ceiling makes it the pick when a scene needs a second reference — an outfit, a prop, a location.
GPT Image 2.6 edit — 5/5. Near-identical through the hardest brief: new pose, new setting, new light. The 16-reference ceiling is the highest in the studio, and the markup canvas (draw on the image to say where to change) is the extra tool when a plain prompt is not precise enough.
FLUX 2 Pro edit — 4/5. The face holds; the wardrobe does not. The collar and strap drifted, which matters when the outfit is part of the character. Still a pass for a story series — just re-check the costume in every frame.
Text only, no reference — 4/5. The control. Works within one model when the sheet is detailed, but the face is approximate and the pose is the model's choice, not yours.

The pattern is consistent with what we found in our Seedream 5.5 review, where edits held a perfume label letter-perfect through a scene swap, and it is the opposite of the MiniMax H3 review, where an edit re-rendered the subject as a close relative instead of preserving her. Not every editor preserves: check the model before you build a series on it.

The workflow: sheet + reference chain

Here is the method, as we would run it for a real series — a comic, a brand mascot, a photo story:

Write the sheet first, generate the anchor. One tight character description, one portrait. Keep the first output you like — that is your reference photo. Generate it at 1K; you are buying a likeness, not a final.
Never regenerate the character from text again. Every later frame is an edit of the anchor, not a new text-to-image call. The control above is what happens when you break this rule: a cousin, not a twin.
Chain edits — but re-anchor. Editing the anchor gives the cleanest likeness. After two or three chained edits the face starts to soften, so every third or fourth frame, edit from the original portrait again.
Describe the scene, repeat the person. Each edit prompt has two halves: "keep the exact same woman — same face, hair, glasses, jacket" and then the new scene. Drop the first half and the model treats the reference as a mood board.
Freeze the frame. Same aspect ratio for the whole series (we used 3:4 throughout, except FLUX at 2:3, which is why its frame is a touch narrower). A series that changes ratio every frame looks like four different shoots.

What each edit costs

Three of the four models charge per extra reference image. Prices below are the live studio rates at 1K (FLUX 2 Pro at 1MP), checked against the manifest before writing:

Seedream 5.5 — 5 credits per edit (4 for the text original), +2 credits per extra reference, up to 10 references. The cheapest edit in this test at $0.02 of provider cost.
Nano Banana 2.1 — 8 credits per edit (7 for text), +1 credit per extra reference, up to 14 references.
GPT Image 2.6 — 8 credits per edit at low quality (5 for text), +3 credits per extra reference, up to 16 references, +3 credits if you use the markup canvas.
FLUX 2 Pro — 10 credits per edit (5 for text), up to 3 references, no per-reference surcharge — the flat edit price is the trade-off for the smallest reference ceiling.

Full pack pricing lives on the pricing page. In pack terms: the whole six-frame test cost 39 credits, under a tenth of the $5 Mini pack, and a ten-frame series — one anchor plus nine Seedream 5.5 edits — costs 49 credits. Consistency is cheap; the expensive mistake is regenerating the character from text every time and throwing away the near-misses.

Where it breaks

Cross-model faces. Every edit above started from a Seedream portrait and stayed faithful — but that is one anchor and four editors. A face generated by one model and edited by another can drift more; if you switch anchor models mid-series, re-check the face frame by frame. Wardrobe drift. FLUX showed it: small costume details (a collar, a strap) are the first things an edit lets go of. If the outfit is the character, name it in every edit prompt. Re-rendering editors. Some models treat a reference as casting direction, not a photo to protect — MiniMax H3's rooftop edit in our review redrew the face outright. Know which kind your editor is before the series depends on it. Multiple people. We tested one person; two characters in one frame doubles the things that can drift, and no model here was asked to keep two faces at once. Text in the scene. Seedream's Kyoto frame spells one sign "Japanese-style Hampers" and renders the rest as plausible-looking Japanese that may or may not be real words — fine for atmosphere, never for a poster that must be read. That job belongs to GPT Image 2.6 or Seedream 5.5 with the copy checked after, per our model guide.

Verdict: the two-line method

Keep the same character in every AI image with two lines: a frozen character sheet, repeated word for word, and a reference photo, attached to every edit. The sheet alone gets you a recognisable character inside one model; the reference is what makes it the same person in every scene and every model. In this test all four editors passed — Seedream 5.5 and Nano Banana 2.1 were near-perfect, GPT Image 2.6 survived the hardest brief, FLUX 2 Pro only let the costume slip. Start with the cheapest edit, Seedream 5.5 at 5 credits, generate your anchor, and build the series from that one frame.

Frequently asked questions

How do I keep the same character in AI-generated images?

Use two things together: a frozen character sheet — the same short description of hair, glasses, clothing and props, repeated word for word in every prompt — and a reference photo. Generate one anchor portrait you like, then make every later image an edit of that portrait instead of a new text-to-image generation. In our test, the sheet alone reproduced the character recognisably inside one model, but only the reference edit kept the exact face across scenes and models.

Which AI model is best at keeping a character consistent?

In our November 13 test, Seedream 5.5, Nano Banana 2.1 and GPT Image 2.6 all scored 5 out of 5 for identity preservation from the same reference photo, and FLUX 2 Pro scored 4 out of 5 — the face held but the jacket drifted. Seedream 5.5 is the cheapest edit at 5 credits; Nano Banana 2.1 takes up to 14 references for multi-prop scenes; GPT Image 2.6 takes up to 16 and adds a markup canvas for precise edits.

Can I keep a character consistent without a reference image?

Partly. A detailed, frozen character sheet reproduced our character recognisably in a new scene with no reference at all — same hair, glasses, freckles, jacket and camera — but the face was approximate and the pose was the model's choice. For illustrations or a loose series that is enough. For a photo story, a comic or a brand mascot where readers must believe it is the same person, use a reference photo.

How much does a consistent-character series cost?

Very little. Our six-frame test — an anchor portrait, a no-reference control and four edits — cost 39 credits in total, under a tenth of the $5 Mini pack. A ten-frame series (one anchor plus nine Seedream 5.5 edits) costs 49 credits. Extra references add 2 credits each on Seedream 5.5, 1 on Nano Banana 2.1 and 3 on GPT Image 2.6; FLUX 2 Pro charges a flat 10 credits per edit with no per-reference fee.

Why does my character's face change between images?

Because each text-to-image call samples a new face that merely matches your description — text describes a type, not a person. The fix is to stop regenerating from text: generate one anchor portrait, then edit that photo for every new scene. Also check the model: some editors re-render the reference as inspiration rather than preserving it — MiniMax H3 did exactly that in our review — so test one edit before committing to a series.

About the author

Written by the Pizyn Studio team. We build and operate this 33-model AI image studio, and every frame in this guide was generated with Seedream 5.5, Nano Banana 2.1, GPT Image 2.6 and FLUX 2 Pro through the same pipeline our users run, with the exact prompts shown and at the prices listed — including the control, which we kept because it taught us something. Prices and model details are checked against the live studio on the date above. Spot an error? Email support@pizyn.com and we will correct it.

Try the models in this guide

Seedream 5.5 · Nano Banana 2.1 · GPT Image 2.6 · FLUX 2 Pro · All models