Keeping a face the same across forty shots

A face that drifts is the single clearest tell in AI film. Not motion, not lighting, not hands. A character whose bone structure shifts half a millimetre between shot four and shot eleven, so the audience stops following the story and starts watching the surface. Most AI brand films fall apart here, and almost none of the published case studies explain what it took to stop it.

Reference capacity is the lever, and it moved an order of magnitude

Everything about consistency comes down to how much evidence you can hand a model about who someone is. That number changed dramatically through 2026. Seedance 2.5 now accepts fifty multimodal references, covering images, audio, 3D white models and style references, against twelve in the previous version. Kling 4.0 takes fifteen references and ten keyframes. Pika H3 Flex takes twelve. MiniMax H3 takes nine images, three videos and three audio references. Luma UNI-1.1 takes nine.

Seedance 2.5

50 refs

Kling 4.0

15 refs

Luma UNI-1.1

9 refs

Capacity is not quality, but it changes what kind of thing a reference set can be. At twelve slots you are describing a look. At fifty you are describing a person: angles, expressions, a voice, a volume in space. That is the difference between a prompt and a casting file.

Two other 2026 releases bear directly on this. Luma's Ray3.2 is the only model we have found publishing performance tracking of skeletal posture, gestures and facial expression for up to eight faces, which is a claim about ensemble scenes rather than single portraits. Runway's Aleph 2.0 allows one edit to propagate across multiple shots and scene changes, which is the first tool-level answer to the complaint that fixing a character means fixing it forty times. On the independent Artificial Analysis board, Seedance 2.5 leads on human anatomy and photorealism, which is the closest thing to a neutral read anyone publishes.

What building one character actually costs

Two documented productions tell you more than any feature list.

Svedka's Super Bowl LX spot took about four months, and most of that went on reconstructing one character, the Fembot, and training the model. One character. Four months. For a national spot with a hard air date. Nobody announcing AI efficiency quotes that figure, and it is the most useful number in the category.

The Swissquote bull had to be rebuilt in 3D with substantial research and development before it could perform physically complex actions and stay consistent and believable. David von Ritter of cavalcade summarised the standard the work was held to: "AI is like plastic surgery: it's good if you can't see it."

Both projects solved consistency the same way, by building an asset outside the model and using generation around it. That is the pattern. When a character has to survive a whole film, somebody builds a thing, and the building is the budget.

If you want third-party confirmation that this is the hard part, read a syllabus instead of a press release. Jason Zada's Secret Level Academy teaches video continuity and consistent characters, and integrating real actors into AI scenes, as named topics. People do not build courses around the easy parts.

The hybrid answer

The most interesting structural response to the problem opened on 2 November 2026. iQIYI's Shanghai AI film studio shoots the performance live, conventionally, and then changes sets, lighting, costume, makeup and props in post. The actor's take is never regenerated. Liang ZHU, who runs intelligent production there, states the reasoning directly: "An actor's performance is the core of any film or series."

“An actor's performance is the core of any film or series.”

Read as a consistency strategy, that design is an admission. The cheapest way to keep a face the same across forty shots is to photograph it once and never ask a model to invent it. Everything generative then happens behind and around a fixed human centre. For a brand film with a human lead, this is almost always the right architecture, and it is cheaper than the alternative because it removes the four-month character build entirely.

Where it still breaks

Three things remain genuinely unsolved in our experience, and we would rather say so than find out on a client's schedule.

• A specific real person. Generating a generic plausible human is reliable. Generating a named individual whose face the audience already knows, holding across a sequence, is not. Likeness is also where the legal exposure lives, which compounds the technical risk.
• A real product in a real hand. The packaging has to be exactly itself, with correct type and finish, held by fingers that occlude it correctly. Two consistency problems colliding at the one point in the frame the client will look at first.
• Physical interaction with weight. Anything where mass has to read correctly: lifting, pouring, pushing, impact. Models produce the appearance of the action without its consequences, and audiences detect the error without being able to name it.

There is a fourth, and it is not technical. In the Ipsos and Syracuse test published in May 2026, fully AI-generated ads performed worst on storytelling, emotion and point of view, and chose a storytelling approach only thirty per cent of the time against forty-nine per cent in Ipsos's human norms. A character who looks identical in every shot and has no interior life is still a failed character. Consistency is the floor of this discipline, not the ceiling.

Who should own the reference set

One commercial point, because it decides whether any of this is worth doing twice. A brand character built properly is a reusable asset. The reference set that keeps it stable, the images, the 3D build, the voice, the trained weights where applicable, is the thing that has the four months of cost in it. That set should belong to the client, named in the contract, deliverable as files at the end of the job.

Studios have an obvious incentive to keep it. A client who cannot regenerate their own character without you is a client who cannot leave. We hand it over, and we would advise asking any studio you brief to put that in writing before the first generation runs.