Guide
Why AI video clips are so short
Five seconds, eight if you are lucky. That ceiling is not a pricing decision or a missing feature — three separate things push against length, and only one of them gets cheaper with time.
Reason one
Every frame has to see every other frame
A video model does not produce frames one after another the way a camera does. It generates the whole clip at once, and the frames have to agree with each other — a hand that moves left in frame ten has to still be moving left in frame eleven, and has to have come from somewhere in frame nine.
Keeping that agreement means each frame is compared against the others. Double the number of frames and you roughly quadruple that comparison work, because every new frame has to be reconciled with every existing one. Length does not cost linearly. It costs something closer to the square.
That is the hard floor under the whole problem. A ten-second clip is not twice a five-second clip, in memory or in time, and the jump is steep enough that most models simply decline to try.
Reason two
The training data was short too
Models learn motion from clips, and the clips they learn from are short — a few seconds each, because that is what is practical to collect, label and feed in at scale. A model trained almost entirely on two-to-five-second fragments has never seen what a thirty-second continuous take looks like.
Ask one for thirty seconds and it does not fail cleanly. It produces something that looks like several short clips stitched end to end, because that is the shape of motion it knows: a gesture completes, then the next thing starts, with no sense that they belong to one continuous action. People read that as the model “losing the plot”, and it is closer to the model never having been shown that plots exist.
Reason three
Drift gets worse toward the end
Small errors accumulate across a sequence. A face shifts by an imperceptible amount per frame, and a hundred and fifty frames later it is a different person. This is the same compounding the video consistency guide covers, and it is the reason length is capped below what the compute alone would allow.
It means the back half of a long generation is the worst half, every time, in a predictable direction. A provider that offers ten seconds is offering five good ones and five that will probably be discarded — which is a bad deal to sell, and part of why the advertised ceilings sit where they do.
So how are long pieces made
Assembled, not generated
Every AI video longer than about ten seconds is a sequence of short generations joined together. There is no exception to this, and any tool offering a “five minute film” is doing it — the question is only whether it does the joining for you or leaves you in an editor.
Which turns the length problem into the continuity problem, and that one has known techniques: seed each shot from the last frame of the one before so the join is clean, keep re-anchoring on the original character reference so the identity does not wander, cut on movement, and keep each shot short enough to stay in the good half.
This is why “how long can it be” is usually the wrong question to ask a tool. The honest answers are how long each individual shot can be before it degrades, and whether the thing assembles shots for you with the character held across them. A ceiling quoted without those two numbers is not telling you much.
Unlimited Studio shoots up to thirty minutes that way: one frame-chained shot at a time, assembled into a single film, with the character reference still applied on each shot. The pipeline page sets out the shot length and how the chaining works.
What to expect next
Which of the three is actually getting better
The compute wall moves. Hardware gets faster and attention gets cheaper to approximate, so the quadratic cost becomes affordable at lengths it was not before. Clip ceilings have risen steadily for exactly this reason and will keep rising.
The training-data wall moves more slowly, because long continuous footage with usable labels is genuinely scarce rather than merely expensive.
Drift is the one that does not obviously yield, because it is a property of generating a sequence rather than a limit of the budget. Expect single-shot ceilings to keep climbing, and expect long-form to keep being assembly for a good while yet.
Keep reading
The rest of the series
- Putting two characters in one sceneIdentity leakage: why one face borrows the other, and why prompt wording is not the fix.
- Why the model ignores half your promptAttention is finite and spread unevenly. Why the end of a long prompt is the part that gets dropped.
- Image-to-video or text-to-video?The choice decides whether you control the first frame or the model does.
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.