Duration is the action's pace, not a length

01How long should a shot be in an AI video?

Time the beat first, then request exactly that. If the action is 2.5 seconds and the model's floor is 4, ask for 4 and accept a slightly slow take. Never round up to be safe. The duration you request is the pace you are directing, and the model has no concept of spare footage at either end.

I learned this the expensive way on a hotel shoot. I sized every shot with a per-shot safety margin: six seconds for a two-to-three second beat, ten seconds for a five-second hold, on the reasoning that a longer window gives you more to choose from. Ninety seconds of generated footage came back for a twenty-five second film, and nearly every take was in slow motion.

You're giving it the action once, and you're giving it seven seconds. So the model is making it really slow, which makes the videos unusable.

Badal Kariwal, reviewing that run

Look at the bottom bar. A quarter speed is a person moving through treacle, and it fails worse than an obviously broken render, because the still frame looks fine. That is the class of error a frame check never catches.

02You are not buying spare footage

A longer take gives you options on a set: the actor does the thing, does it again, you pick. A generator does not work that way. It will not perform the action three times inside seven seconds so you can choose the good 2.3. It stretches the one action to fill.

That last row is a rule in this studio and it is worth stating on its own, because it changes how you budget: a bad generated take is re-rolled, never repaired. Speed-ramping a slow take back up strobes it. Cutting the middle out of it produces a jump in a shot that had no cut in it. The take is the unit.

03How do you get more material inside a take?

Ask for scene cuts inside the take rather than a longer single action. One call, three or four internal scenes, each with its own timeline slice written out in the prompt, each grouped by place so every segment stays anchored to something real. You get four usable beats at natural pace out of one window instead of one unusable beat stretched across it.

ONE CALL, 10s, one location, three scenes

[0-3.5s]  hands enter frame, lift the jar off the shelf
[3.5-6.5s] quarter turn, label comes round to camera
[6.5-10s]  set down on the counter, hands leave frame

rules that make it work:
  every slice is a real action at its own natural speed
  every slice stays in ONE place, so nothing has to be invented
  no slice is longer than the action it names
  the word slow appears nowhere unless slowness is the point

Slices work because they take back a decision the model was making on your behalf. A ten-second window with one action in it is under-specified, and the model resolves it the only way it can, by slowing down. Three timed actions in the same window leave nothing for it to decide.

04A tempo correction that saves one model ruins another

Generated motion often reads slow at native speed, so it is normal to nudge a clip faster before it enters a cut. We ran a standing 1.1 times correction on everything. It was right for the models it was written against, the ones that render people moving slightly underwater. On a model without that problem, the same nudge makes hands twitch.

Write the number down, with the model it was measured on

So the correction became a per-model number instead of a house constant. Measure the model's native pace on a clip whose real duration you know, apply a correction only where the footage genuinely reads slow, and write down what you applied in the build notes. That last part sounds like admin and is the part that matters: an undocumented multiplier surfaces the day somebody swaps models, when nobody remembers it exists.

05When does a long take beat a cut?

When the shot has an arc of its own: something starts and finishes inside the frame, or the take carries its own scene cuts. That is the only test, and you apply it to the shot itself, watched cold at 360 pixels, which is the real width of a feed tile. Not to its description in a document.

Liquid, hair, fabric and breath price a mistake instantly, because each carries its own physical clock. If one of them is in your beat, pace is not a taste question. They are also the events current models break most often, and that is not a coincidence: the clock that makes a pour readable is what makes it hard to fake.

06Two shot-length shapes that both beat an even rhythm

Even is the enemy. Twenty rigorous films built to one cut rate read as a single film, and that sameness is what a viewer notices first and a maker notices last. Two shapes beat it, and they are opposites. One accelerates. One decelerates. Both are a direction, which an even rhythm never is.

The acceleration shape runs the other way: three cuts across six seconds to set a pulse, then four cuts of 8, 14, 11 and 9 frames inside 1.4 seconds. And there is a third move worth stealing, for a sequence of similar shots that would otherwise read as an assortment. Halve the cut rate exactly halfway through: 45, 30, 30, 30, then 15, 15, 15, 15. The shots become a series because the rhythm is going somewhere.

Every one of these mistakes is a margin somebody added to protect themselves. Longer windows, safer durations, a correction applied to everything so nothing gets missed. The margin always moves the cost from the person adding it to the film.