AI Video Character Consistency: How to Keep One Character Across a Whole Campaign

In the hook, your spokesperson has a small scar over her left eyebrow. By the third ad it's gone. By the ninth her jaw looks softer, her jacket is a different green, and someone in the comments asks if you hired a new actor.

Nobody planned that. It happened because the campaign assumed the generator would keep the character consistent, and generators don't do that on their own.

This guide is for brands commissioning AI video ads from a creator. It covers how good creators keep a character stable across a dozen or a hundred assets. It also covers what you, the buyer, should ask for and approve, so you find drift before the media budget is spent and not after.

Why the usual advice falls short

The standard advice on AI video generation character consistency is a familiar list: reuse the prompt, fix the seed, upload a reference image, and regenerate until the face looks close enough.

Those tactics help with a single shot. They won't hold a campaign together. A repeated prompt and a fixed seed can still return a different hairline once the camera angle, lighting, or scene changes. "Close enough" also ends up as a shot-by-shot judgment, made by whoever is tired at 11pm, with no written standard to check against.

The real fix is less exciting than a prompt trick. You need an approved source of truth, meaning a reference set plus a written list of what must not change. Then you need a review process that checks every deliverable against it. Consistency is an asset and workflow problem. Once you treat it that way, much of the drift either goes away or gets caught before it ships.

The five layers of continuity

A campaign doesn't read as one system just because the face matches. These are the layers that have to hold:

The first two are easy to confuse, but they're different problems. A clip can be completely stable within itself, with no flicker and no melting during movement, and still show a different person from the clip before it. The reverse happens too: a face that matches the reference in stills and then morphs mid-turn. This technical breakdown of cross-shot vs. temporal consistency explains why each one needs its own fix.

For you, that means checking both. A creator who only reviews thumbnails can easily miss temporal drift, because it shows up in motion, not in stills.

The lock / vary / verify workflow

This AI video character consistency workflow holds up across a full campaign. It has three stages, and the order matters.

Lock: build and approve the character pack

Before anyone generates motion, the creator builds a character pack and you approve it. Every asset that follows is checked against this pack.

A solid pack usually includes:

• An AI character reference sheet showing the character from the front, three-quarter, and profile in neutral lighting, plus at least one full-body view
• Close-ups of distinguishing details such as moles, scars, freckles, tattoos, eye color, and teeth if the character smiles on camera
• Wardrobe reference with exact colors and how the logo sits on the garment (size, position, which side)
• Reference images for any prop the character holds or wears more than once
• A voice reference with performance notes on pitch, pace, accent, energy, and how the character gestures
• A written invariants list: everything that must not change in any asset

The invariants list is the easiest part to skip and the most useful one to have. Reference images are open to interpretation. A sentence like "Gold hoop in the left ear only. Logo on the right chest, never mirrored. Hair parted on the left. Measured speaking pace" is not.

Next, approve keyframes before any motion work. Ask the creator for a handful of representative stills: the hook frame, a product moment, a close-up, and a wide shot. Approving stills is cheap. If you reject finished video because the character drifted, you pay for generation, editing, and voice work twice.

Vary: change one layer at a time

With the character locked, production of variants starts. This is where campaigns tend to lose control, because everyone wants to test everything at once.

Change things in this order:

1. Copy and script first, with scene and character fixed
2. Then the scene or setting
3. Then camera angle and motion

Suppose scene, wardrobe, camera, performance, and script all change in one batch and the character drifts. Nobody can tell which change caused it. The same thing happens to your ad results: if variant B beats variant A and five things changed between them, you've learned close to nothing.

The decision rule: if a change does not serve the test hypothesis, lock it.

Testing a new hook? Lock the scene, wardrobe, and camera. Testing a kitchen setting against an office? Lock the script and wardrobe. Anything left free to wander adds noise to your test data and gives the character another chance to drift.

Verify: check every deliverable against agreed tolerances

Verification means comparing every final asset side by side with the approved pack. A quick look at the hero ad doesn't count.

A proper verify pass includes:

• Freeze-frame comparisons of the face against the reference sheet at similar angles
• Scrubbing every head turn, hand-to-face movement, and fast motion at reduced speed, since that's where temporal drift hides
• Going through the invariants list item by item, including small things like earring side and logo orientation
• Playing voice tracks from different variants back to back to catch pitch and pacing changes
• Checking every aspect ratio export, not only the master

"Agreed tolerances" means you decide in advance what's allowed to move. Lighting can change with the scene, and expression changes with the script. The jawline, the scar, and the logo position stay fixed.

Cross-format: where crops break continuity without anyone noticing

Many campaigns ship in several shapes: 9:16 for vertical placements, 1:1 for feeds, and 16:9 for horizontal video. Each crop is another place for continuity to break.

• A signature prop held at waist height is in frame at 16:9 and gone in a tight 9:16 crop.
• In vertical, a chest logo can end up under captions or on-screen interface elements.
• Flipping a shot horizontally to fix composition moves the earring, the hair part, and the logo to the wrong side. Mirroring is a quick layout fix, and the side swap is easy to miss in review.
• Generating a shot natively in a new aspect ratio, instead of reframing an existing one, creates a new generation. Its identity has to be verified again from scratch.

Put the full list of target ratios in the brief, and ask the creator to frame for the tightest crop from the keyframe stage. Platform specs matter less here than making sure the character's identifying details survive the smallest frame you'll run.

Common failure modes and how to fix them

1. The face matches in stills but morphs mid-turn. Thumbnail review won't catch this. Review motion at reduced speed. Then regenerate the segment from approved keyframes, shorten the movement, or cut on the turn.
2. The jacket or logo changes between cuts. Write exact wardrobe and logo placement into the invariants list. If the generator keeps reinterpreting the logo, the creator should composite it in post instead of regenerating and hoping it lands.
3. An earring swaps sides after a mirrored shot. For any character with asymmetric details, ban mirroring outright or list each asymmetric detail so the editor knows to check it.
4. Voice or timing drifts between variants. Use one approved voice reference with pacing notes, produce the voice for every variant under the same settings, and compare tracks back to back before delivery.
5. A vertical crop cuts off the signature prop. Frame for the tightest ratio at the keyframe stage, not during the final export.
6. Several variables changed at once and nobody knows what broke. Go back to the last approved version and reintroduce the changes one at a time. It feels slow, but it's faster than guessing.

When perfect consistency isn't worth paying for

Consistency costs time. Every extra round of generation, compositing, and review shows up in the timeline and the invoice, so spend it where viewers will notice.

Tight consistency is mandatory for:

• A recurring spokesperson or brand character who appears across many ads
• Multi-market campaigns where the same character appears in localized versions
• Sequential or episodic ads where the same viewer sees the character again and again
• Ads where the character is effectively the face of the brand

It's usually not worth the cost for:

• Background characters who appear in a single shot
• Early concept tests that only check whether an angle works at all (lock the character later if the concept wins)
• Crowd scenes, hands-only shots, and anyone out of focus

The rule of thumb: the more often one viewer sees the character, the tighter the tolerance has to be. A barista who's on screen for two seconds in one ad can be roughly the same person. A spokesperson who appears in all twelve ads can't.

What buyers should ask for and approve

This is where many consistency problems are prevented or created. If you only approve finished videos, you're reviewing too late. It's the same trap behind why one-off AI videos fail: each asset gets judged alone instead of as part of a system.

Deliverables to request

Put these in the brief or statement of work:

• Character reference sheet with multiple angles, neutral lighting, and a full-body view
• Detail close-ups of every identifying mark
• Wardrobe and prop reference, including exact logo placement
• Written invariants list
• Voice reference and performance notes
• Approved keyframes for each scene before motion starts
• A variation plan showing which layer changes in which batch
• Exports for every target aspect ratio
• A provenance and consent record (covered below)
• A clear agreement on who keeps the character pack and source files afterward

If you need a starting template for the brief, this guide on how to brief an AI video creator covers the rest of it.

Acceptance table

Use this as the review standard. Better still, attach it to the contract so "consistent" means the same thing to both sides.

Questions to ask before you hire

• How do you keep a character consistent across separately generated shots? Strong answers mention a reference pack, invariants, and keyframe approval. "Same prompt and seed" is a weak answer.
• Can I see a past campaign where one character appears in multiple ads?
• Do you review motion at reduced speed, and what are you looking for?
• How do you handle logos and on-screen text?
• If I find drift after delivery, does the fix count as one of my revisions?
• Which aspect ratios do you frame for from the start?

For a broader look at sourcing and vetting, see how to hire an AI video creator. Viralix is one place to find vetted AI video creators who produce campaign-ready assets. Whoever you hire, bring these questions to the first call.

Consent, usage scope, and provenance

These belong in the character pack as operational questions, not as an afterthought for legal. None of this is legal advice. It's the list of things you should be able to answer before a recurring character goes live.

• Is the character wholly synthetic, or is it based on a real person's face, body, or voice?
• If a real person is involved, who approved the use of their likeness and voice, and where is that consent documented?
• What's the permitted use: which campaign, which media, which territories, and for how long?
• Can the character pack be reused in future campaigns? Can the creator reuse the character for other clients?
• What happens to the source assets when the engagement ends?

If a professional performer's likeness or voice is involved, union terms may apply. SAG-AFTRA's guidance on AI in commercials shows how a performers' union approaches digital replica consent and use. It governs that union's contracts, not every production, but it's a good model for the questions to ask.

On provenance, C2PA Content Credentials attach tamper-evident metadata that records how a file was made and edited. They're a useful part of your audit trail. They don't prove the ad is truthful, and they don't prove the rights are cleared. Keep them alongside the consent record and usage terms, not in place of them. For more on who owns what, read the AI video rights and licensing guide.

Before your next brief goes out

1. Add a written invariants list to the brief. Ten specific lines about what must never change will usually do more to prevent drift than another prompt technique.
2. Approve the character pack and representative keyframes before any motion is generated, and say in writing that this is the gate.
3. Attach the acceptance table to the statement of work, so every deliverable is judged against the same standard, in every aspect ratio, at reduced speed.