Why it is harder than with stills
With a still you have to get one face right, once. With video you have to get the same face right across hundreds of frames, while the head turns, the light changes and the character speaks. Every frame is another chance to drift.
And that is the easier case. The harder one is sameness across cuts: when the character is in a wide shot in one take, in a close-up in the next, and comes back in the third. To the model these are separate generations — nothing ties them together except what you supply.
The four inputs you can work with
- A reference image. The most widespread one: a photo or several of the character. The more angles the better — but the allowance is tight, and that is the next section.
- A reference video. A newer option, and a stronger one: the model picks up not only the face but the movement. At one provider the voice comes across too from the reference video, so the character's timbre stays stable as well.
- A first frame for every shot. The most underrated method. If you make the opening frame of each setup in an image generator and settle identity there, the video model only has to move it.
- Style and environment reference. This pins down not the character but the world — light, palette, location. Without it the character is right and the scene is not.
How much the model accepts
The allowances differ enormously from provider to provider, and that on its own is a tool-selection criterion.
- The most generous model takes thirty images, ten videos and ten audio samples for a single generation — fifty references in total.
- The middle of the field sits around nine images, three videos, three audio clips.
- At several large providers, though, the ceiling is three images and nothing else.
- And there are ones where it is four — and not mixed there either, images only.
If you are seriously producing a series with a recurring character, that number weighs more than a leaderboard position.
Where the benchmark sobers you up
In 2026 a benchmark built specifically for this appeared, examining sameness across cuts. Its main finding is that identity preservation degrades steeply with return distance: the more shots pass before the character shows up again, the less the same they are.
Translated into practice: a character who is on screen throughout is still fine; a character who appears in the first shot and in the eighth is not. If the story calls for that, the return has to be forced not out of the model but out of the first frames.
The trick worth taking from the benchmark
The same benchmark has an even more important, methodological finding. The earlier metrics simply looked at how much the character resembles itself from frame to frame. The trouble with that is that the model scores high even when it draws the same consistently wrong face.
So they introduced a separate check on whether the character is the right character at all. For you the lesson is the same: do not look at whether the eight shots resemble each other, look at whether they resemble the original reference. Those two are not one and the same, and drift slips in through exactly that gap.
The workflow
1. ONE reference set, and never depart from it.
3-5 images: front, three-quarter, profile, one full body.
Clean background, even light, neutral expression.
2. A FIRST FRAME FOR EVERY SHOT, out of that same set,
made in an image generator. Identity is settled here, not in the video.
3. THE VIDEO PROMPT should not describe the character again.
It is already in the first frame. The prompt is about movement.
4. COMPARE EVERY FINISHED SHOT against the ORIGINAL reference,
not against the previous one. At the end of a chain you end up elsewhere.
5. ON A LONG RETURN (5+ shots) hide the character or take them
out of frame, and bring them back in a close-up.
Common misconceptions
- "If the reference is good, the model handles it." Inside one clip, largely yes. Between shots, no — there the first frame is the answer.
- "More references are always better." Near the upper limit it gets worse if the images contradict each other. Few but consistent references beat many mixed ones.
- "If the shots resemble each other, the series is fine." This was a trap in the benchmarks too: a consistently wrong face is also "consistent".
- "One photo is enough." From one the model guesses the other views. Three to five angles make an order-of-magnitude difference.
What to take away
Character identity in video is not the video model's job, it is the first frames' job. One fixed reference set, a first frame made in an image generator for every shot, and every check made against the original — those three steps give you a stable character. Which reference type carries what across is what Reference Images goes into, and how the shots add up to a finished film is covered by Building a One-Minute Video, the next post along.