Guide
Keeping a character consistent in AI video
Video drifts twice. Once inside a clip, as the frames are interpolated, and again at every cut, because the next clip is a fresh generation that remembers nothing.
Two different problems
Drift within a shot, and drift between shots
People describe both as “the character changes”, and they have different causes and different fixes, so it is worth separating them before trying anything.
Within a clip, the model is producing a sequence that has to stay coherent frame to frame. It is good at that over a second or two and gets worse the longer it runs. Faces are where you notice first, because a face is the thing humans read most precisely — a drift small enough to be invisible on a hand is a different person on a face. Hair partings swap sides, jaw width wanders, an earring vanishes.
Between clips, nothing drifts at all. The second clip is simply a new generation, which is the same problem the first guide describes, now with motion attached. It has no memory of the clip before it unless you give it one.
Technique one
Frame chaining, and what it costs
The obvious fix for the cut is to take the last frame of one shot and use it as the opening frame of the next. Continuity is then near-perfect across the join, because the second shot literally begins on the first shot’s last image.
It has a cost that shows up slowly. Each shot inherits whatever the previous one drifted to, and then adds its own drift on top. Five shots in, you are conditioning on an image that is already four generations away from the character you created, and the compounding is invisible from any single cut. It only shows when you put shot one beside shot five.
Chaining also forces the shoot to be sequential — shot five cannot start until shot four exists, because its opening frame does not exist yet. That is a real constraint on how fast a sequence can be produced, and it is why a film assembles over minutes rather than all at once.
Technique two
Re-anchoring on the original reference
The counter-move is to condition every shot on the original character reference rather than on the shot before it. Nothing compounds, because every shot is measured against the same fixed thing, and shot five is no further from the character than shot one.
What you lose is the join. Two shots anchored to the same reference will agree about the face and disagree about everything the reference does not specify — exactly where the hair is sitting, how the shirt is creased, which way the light falls. Cut between them and it reads as a continuity error.
The two techniques fail in opposite directions, which is the useful part: chaining preserves the join and loses the character, anchoring preserves the character and loses the join. Any system that holds a character across a real sequence is doing some of both.
In practice
What to do with this
Keep shots short. Most of the within-clip drift arrives in the back half of a long generation, so two shots of five seconds hold a face better than one of ten, and you get a cut out of it.
Cut on movement rather than on a still face. A viewer forgives a great deal across a cut that lands mid-gesture, and forgives almost nothing across a cut between two near-identical head-and-shoulders framings, where the small differences are all there is to look at.
Check the ends, not the middle. Drift is monotonic within a shot, so the last frame is the worst frame — and under chaining it is also the one the next shot inherits. If the last frame is wrong, re-run that shot before shooting the next one, because everything after it will be built on the mistake.
Unlimited Studio shoots a film one frame-chained shot at a time, with the character reference still in play on each shot, which is the mixture the section above describes. The pipeline page sets out exactly where each one applies.
Keep reading
The rest of the series
- Why the outfit and the room change tooWardrobe and location drift for a different reason than faces do, and need a different fix.
- Why AI video clips are so shortThree separate reasons for the ceiling, and why longer films are assembled rather than generated.
- Putting two characters in one sceneIdentity leakage: why one face borrows the other, and why prompt wording is not the fix.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.