How to Hold the Same Face Across an AI Campaign

A campaign is not a collection of good images. It is the same world, seen more than once. The moment the face changes between frame three and frame seven, the viewer stops reading a campaign and starts reading a mood board, and no amount of individual frame quality repairs that.

This is the problem that separates people making pretty AI pictures from people producing campaigns. It is also the problem most guides avoid, because the honest answer begins with a limitation: you cannot solve it by writing a better prompt.

Why Prompting Cannot Solve It

A prompt describes a type of person. “A woman in her early thirties, strong brow, wide-set eyes, close-cropped dark hair” is a genuinely specific description, and it will still return a different person every time you run it.

That is not a failure of your writing. It is what the description does. You have narrowed the model from every possible face to a family of faces that satisfy the description. Across twenty frames you will get twenty members of that family. They will look related. They will not look like one person, and viewers tend to notice, even when they cannot say what is wrong.

So the first discipline is to stop trying. Once you accept that identity must be carried by something other than language, the methods below become obvious rather than clever.

Method One: Reference Stacking

The most accessible method: hand the model images of the face you want and instruct it to carry identity from those.

The mistake almost everyone makes is supplying the wrong references. Directors tend to attach their three favourite frames, which usually means three flattering portraits, similar angle, similar light, similar expression. That set is redundant. It tells the model the same thing three times and leaves it guessing about everything else.

A useful reference set varies angle and holds everything else steady:

Neutral light matters more than people expect. A reference shot with dramatic side lighting teaches the model that the shadow is part of the face. You will then fight that shadow in every frame you generate afterwards.

Where the ceiling sits

There is no universal number, because the limit belongs to the model. Google documents different figures for different Gemini image models: gemini-2.5-flash-image works best with up to three input images, while gemini-3-pro-image supports five images with high fidelity and up to fourteen in total. Black Forest Labs documents up to eight references through the API for FLUX.2 editing, and up to ten in its playground. Both pages were read on 27 October 2026 and provider limits change, so check the current documentation before you plan a shoot around a figure.

Our own rule of thumb, which is a judgement and not a measurement, is that one to three references is enough for look consistency, and that adding more mainly adds contradiction. For identity, choose a model built for multi-reference input and supply several angles of the same face under neutral light, rather than choosing for single-frame beauty.

Method Two: Identity Models

A step beyond references: train or register a persistent identity that you then call by name across every generation. Different platforms name this differently, but the principle is constant, the face stops being an input you re-supply each time and becomes an asset you reference.

In our judgement this is usually the strongest single method, although it depends on the platform and we have not measured it against the others. It comes with the obligations you would expect. An identity you can reuse indefinitely is, functionally, a model on a permanent contract. Treat it that way.

The permissions question

If the identity is built from a real person, you need their informed, written consent for the specific use, and it should name AI generation explicitly rather than hiding inside a standard usage clause. A model release written for a photoshoot in 2019 did not contemplate a permanent synthetic likeness, and pretending otherwise is both an ethical failure and a commercial risk sitting inside your campaign.

If the identity is wholly synthetic, disclose that where it matters. Disclosure is also becoming a legal question. In the EU, Article 50 of the AI Act requires disclosure that image, audio or video content resembling real people, places or events has been artificially generated or manipulated, with a limited duty where the work is evidently artistic or creative. Whether and from when it applies to your campaign is a question for your legal adviser, and this article is not legal advice.

Method Three: Seed Discipline

Seeds are widely misunderstood. A fixed seed does not mean a fixed face. It means a repeatable starting point, which produces the same result only when everything else is identical too. The Hugging Face Diffusers documentation is plain that diffusion is a random process, that repeating a result depends on how the random generator is set up, and that a generator on a GPU can diverge from one on a CPU. Where a hosted model exposes a seed at all, check what its provider says the seed guarantees.

Change the prompt meaningfully, change the aspect ratio, or change the model version, and the same seed gives you a different person. Seed discipline is not an identity method. It is a variation method, and it is excellent at that job.

The correct use: once you have an approved frame, hold the seed and change one variable at a time to explore around it. Same seed, same subject text, new lighting sentence. You get a controlled family of options rather than twenty unrelated attempts. Treat this as a technique to test on your own model, not as a guarantee.

Method Four: Edit, Do Not Regenerate

This is the method that changes campaigns, and the one most teams arrive at last.

Regeneration re-decides everything. Every time you run a prompt again, the model makes fresh choices about the face, the hands, the fall of the fabric, the light, including all the choices you had already approved. You are rolling the dice on work you had already won.

Prompt-driven editing is meant to do the opposite. You take an approved frame and change only what you name. Different background, same subject. Different garment, same face. Different time of day, everything else held. Editing models can still alter areas you did not name, so lay each result beside the approved frame and check the face, hands and fabric before it goes into the set.

How a campaign is actually built

This inverts the instinct most people bring, which is to generate twenty frames and pick the best five. That approach makes inconsistency likely, because the five you pick were each produced by an independent roll. Building outward from one approved frame makes consistency a matter of method rather than luck.

Consistency in Motion

Video is harder, for a straightforward reason: the model must hold the face across many frames while moving it, and small errors accumulate along the clip. A face that is stable at second one can drift by second eight.

Three practical responses, in order of effectiveness:

Give it a start frame

Never describe your subject to a video model in text when you can hand it your approved still as the clip's first frame. In our judgement this is the most useful single step, and it is why an approved hero frame is worth so much: it seeds the stills and the film.

Keep clips short

Longer clips give error more time to accumulate. Four to six seconds is our working rule, not a measured threshold, and the right length differs by model. Cut two short clips together rather than asking for one long one.

Design shots that do not depend on the face

The oldest trick in film direction and still the best. Shoot from behind. Hold on hands, on fabric, on the turn of a coat. Let the face appear briefly and in a held frame rather than through a long movement. These are legitimate directorial choices, not workarounds, a great deal of fashion film has always been shot this way.

The Continuity Sheet

Whichever methods you use, write the identity down. A campaign that runs across stills, film and social needs a single reference document that anyone touching it can follow, exactly as a physical production would keep continuity notes.

The value of this sheet is that it makes the distinction explicit: what is identity and what is styling. Most drift happens because someone changed a thing nobody had written down as fixed.

The Four Methods Compared

A summary of the sections above. The last column says how each row is backed: provider documentation, or our professional judgement.

What This Article Does and Does Not Show

This article was written and reviewed by House of Essenzi as an organisation. No individual reviewer is named, and none is implied. The advice is professional judgement about how campaigns are built, checked against the provider documentation listed below on 27 October 2026.

We have not published a controlled comparison for this article. There is no measured failure rate here, and none should be inferred from the wording. Where a sentence says “usually” or “in our judgement”, it is a judgement. Where it gives a provider limit, it is taken from the source named beside it.

Run your own test

Because providers change their models, the most reliable evidence is your own. A fair test is small enough to repeat:

Report the model, version, date, input method, settings and sample size alongside any result you share, so that someone else can repeat it or see why they cannot.

Frequently Asked

Sources

Read on 27 October 2026. Provider documentation changes, so the current page may differ.

1. Image generation with Gemini (Nano Banana), Google AI for Developers. Reference limits differ by model: gemini-2.5-flash-image works best with up to 3 input images, gemini-3-pro-image supports 5 images with high fidelity and up to 14 in total, and gemini-nano-banana-2.1 and gemini-3.1-flash-image support character resemblance for up to 4 characters.
2. FLUX.2 Image Editing, Black Forest Labs documentation. FLUX.2 editing takes multiple reference images: up to 8 through the API and up to 10 in the playground.
3. Reproducibility, Hugging Face Diffusers documentation. Diffusion is a random process. Repeating a result depends on how the random generator is set up, and a GPU generator can diverge from a CPU one, so a seed alone does not guarantee the same output across machines.
4. Regulation (EU) 2024/1689, Article 50 (transparency obligations), EUR-Lex, Official Journal of the European Union. Article 50 requires disclosure that image, audio or video content resembling real people, places or events has been artificially generated or manipulated, with a limited duty where the work is evidently artistic or creative.

Corrections and revisions

• 27 October 2026. A locked seed was listed alongside images and identity models as a way to carry identity. That contradicted the section on seeds, which says a seed is a variation tool. It now appears only as a way to repeat a result, in the answer, the body and the structured data alike.
• 27 October 2026. The claim that most models take one to three references was replaced with the limits each provider documents, dated, with our one-to-three figure labelled as a rule of thumb. Sources, a method comparison and a way to run your own test were added.