How to repeat a camera move across takes
Slow dolly is re-estimated on every generation. Parameter-based camera control and endpoint framing are the two ways to lock one move across a whole sequence.
Write slow dolly in six times for six shots in a sequence and you get six different moves. One arrives in two seconds and sits. One creeps for the whole clip and never finishes. One starts, hesitates, and drifts sideways. Every take is defensible in isolation and the sequence is unusable, because a signature move that changes every time isn't a signature — it's noise the audience reads as inconsistency.
This is a different problem from a camera instruction being ignored. The instruction landed. The model dollied. What it didn't do is dolly the same amount at the same rate as last time, and nothing in the prompt asked it to, because prose has no way to carry a velocity.
Why the move changes every take
A camera move has at least four properties: direction, distance, speed, and easing — how the move accelerates in and decelerates out. Text carries the first one well, the second one roughly, and the last two not at all.
"Slow" has no unit. It is relative to the distribution of moves the model has seen, and where any individual sample lands in that distribution is decided at generation time. So is easing: whether the push ramps gently or snaps to speed in the first ten frames is being re-estimated from scratch on every run. Two takes with byte-identical prompts can differ by a factor of two in how far the camera travels, and that difference is not a bug being fixed in the next model version — it is the consequence of describing a continuous quantity with an adjective.
Turning up a motion-level dial doesn't help either. Where a model exposes one, it is a coarse "how much changes per frame" control across the whole image, not a camera-speed selector, so raising it adds movement to the subject and background alongside the camera rather than making a described push more decisive.
Two things do lock the move. One is turning it into a parameter. The other is fixing where the shot starts and ends and letting the middle be approximate.
Lock 1: parameter-based camera control
Some models take camera moves as configuration rather than prose. Where that exists, the move stops being interpreted: the same value produces the same shape of move on every take, in the same way that setting a duration produces the same duration.
This is a narrow slice of any large catalog. Of the 146 active video models on Versely, roughly one in ten carries an explicit camera control capability rather than leaving camera moves to be inferred from ordinary prompt vocabulary — in practice the Seedance 2.0 family, the Kling V3 and O3 Pro lines, and Runway Gen-4.5. A smaller set again is built specifically around transferring a move, including Kling V3 Pro motion control; the current lineup for that is rounded up on the best motion control model page.
The practical consequence for a sequence is worth stating plainly: if repeatability matters more than any other property of the shot, pick the model for its camera control and accept whatever else it gives you. That inverts the usual selection order, where you choose on look and quality first. It is the right inversion when the move is the through-line of the piece — a product line where every item gets the identical orbit, a title sequence where each card gets the same push, a six-part series that has to feel like one crew shot it.
Where the capability doesn't exist, no amount of prompt engineering substitutes. That is not a phrasing problem to solve; it is a missing input.
Lock 2: endpoint framing
The second lock works everywhere and constrains something slightly different. Instead of specifying the move, you specify the two frames the move has to connect.
First and last frame generation takes both endpoints as images and interpolates between them. The path in the middle is still the model's decision, but the start framing and the end framing are yours, which means the magnitude of the move is fixed even when its easing isn't. A push that begins with the subject at 40% of frame height and ends at 70% is the same push every time you supply those two frames, regardless of how the model chooses to get there.
For a repeating move across a sequence, this turns into a mechanical recipe:
- Build the start frame for shot one as a still, in whatever image model you trust for the look.
- Build the end frame by cropping the start frame to the framing the move should finish on, then re-minting that crop as a clean image rather than feeding a stretched crop forward.
- Generate the clip from that pair.
- For every subsequent shot, apply the same crop ratio to that shot's start frame to derive its end frame.
Step four is the whole trick. Because the crop ratio is a number you control, every shot in the sequence travels the same proportional distance, and the sequence reads as one camera behaving consistently. Models that take both endpoints include Veo's first-last-frame mode and Flux 3's equivalent.
The cost is real and worth naming. Endpoint framing removes the model's freedom to compose the end of the shot, so anything you would have wanted it to discover — a nice settle, an unplanned reframe — is gone. On a hero shot that's a loss. On shot four of six in a matched sequence, it isn't.
What doesn't lock it
Three things people reach for that don't solve this:
Seeds. A fixed seed reproduces one generation from one prompt. Change the prompt — and every shot in your sequence is a different prompt — and the seed no longer refers to the same thing. Seeds are for getting a result back, not for carrying a move between shots.
Intensity adjectives. "Slow", "gentle", "smooth", "subtle" modulate but do not specify. Stacking them ("very slow, gentle, barely perceptible dolly") narrows the distribution slightly and widens the chance of a contradiction, because "barely perceptible" and "dolly" pull against each other and the model resolves the conflict by picking one.
Referring back. "The same camera move as the previous shot" refers to something the model cannot see. Each generation is independent; there is no previous shot in its context.
One more thing worth getting right while you're here: name the mechanism, not the effect. "Closer" is satisfied by a dolly or a zoom, and those look completely different — a dolly changes perspective as the camera physically travels, a zoom flattens it. Write the camera physically travels toward the subject when you mean a dolly. If a text description isn't holding at all, handing the model a motion reference clip is the third option, and the trade-offs of that route are in describe the camera or hand it a reference clip.
Measuring whether the move repeated
Do not judge this by eye across six clips played in sequence. Measure it, on stills, in a couple of minutes.
For each take, grab the first and last frame and measure one thing in both — the subject's height as a fraction of frame height works for pushes and pulls, the subject's horizontal position works for pans and tracks. Two numbers per take. The ratio between them is the magnitude of the move.
| Take | Subject height, frame 1 | Subject height, last frame | Ratio |
|---|---|---|---|
| 1 | 0.40 | 0.70 | 1.75 |
| 2 | 0.41 | 0.55 | 1.34 |
| 3 | 0.40 | 0.82 | 2.05 |
That spread is what "slow dolly in" produced three times, and it is immediately visible as a number while being genuinely hard to spot by watching. Set a tolerance — ratios within about 10% of each other cut together comfortably — and reroll or re-crop anything outside it.
Easing won't show up in this measurement, and that is fine. Magnitude mismatch is what audiences notice across a cut; easing differences of the same magnitude mostly read as natural variation between takes.
FAQ
Why does the same camera prompt produce a different move every time?
Because speed and easing are continuous quantities and prose has no unit for either. "Slow" positions the move somewhere in a distribution the model learned, and each generation samples that distribution fresh. Direction survives text well, distance survives it roughly, and rate does not survive it at all — which is why identical prompts return moves that differ in how far and how fast the camera travelled.
Does a fixed seed make a camera move repeatable?
Only for the exact same prompt. A seed reproduces a specific generation; it does not transfer a camera behaviour to a different prompt. Since every shot in a sequence has different subject and setting text, the seed is anchoring a different result each time. Use seeds to get one good take back, not to standardise a move across a set.
Can I get a repeatable move on a model that has no camera-control parameter?
Yes, through endpoint framing. Supply the start and end frames and the magnitude of the move is fixed regardless of how the model handles the middle. Derive each shot's end frame by applying the same crop ratio to its start frame and the whole sequence travels the same proportional distance. You give up control of the easing and any freedom the model had to compose the end of the shot.
Is it better to pick a model for its camera control or its image quality?
It depends on whether the move is the point. For a single hero shot, pick on quality — you can iterate on the move. For a sequence where a repeated move is the through-line, pick on camera control first, because a look mismatch can often be reconciled with a colour pass at assembly and a move mismatch cannot be fixed after the fact at all.