Guides

    How to get a genuinely locked-off shot

    Leaving camera language out of a video prompt does not give you a static frame. The explicit phrasing and staging that produce a truly motionless camera.

    Versely Team8 min read

    The intuition is reasonable and it is wrong: if you want a static camera, say nothing about the camera. No pan, no dolly, no orbit, so nothing moves. What you get back instead is a slow drift, a gentle push that arrives from nowhere, or a handheld float that nobody asked for.

    Models carry a strong prior toward adding movement. Their training data is video, and video that sits perfectly still is a minority of it. Absence of a camera instruction is not read as "hold still" — it is read as "unspecified," and unspecified gets filled with the average of everything the model has seen, which moves. If you want a locked-off frame you have to ask for one, in words, explicitly.

    Say the stillness out loud

    The phrasing that works is a direct statement of the camera's state, not the omission of a movement verb. Some versions that hold up:

    • The camera is entirely motionless.
    • Locked-off shot on a tripod. The camera does not move at any point.
    • Static camera, fixed frame, no pan, no tilt, no zoom, no drift.

    Two things are doing the work there. The first is that stillness is stated as a property of the camera rather than as an absence. The second is redundancy — naming the specific moves you are excluding as part of a positive description of a fixed frame, rather than as an instruction not to do them.

    That distinction matters more than it sounds. Instructive negation ("do not pan," "no camera movement") tends to leak the negated concept back into the frame, so the reliable move is to describe the desired state instead: not "no buildings," but "a desolate landscape with no buildings or roads." The second phrasing still contains the word "buildings" and still works, because it is a description of a scene rather than a command about what to avoid. Apply the same shape to camera: describe a tripod-locked frame that includes the fact of its stillness, rather than issuing a list of prohibitions.

    This is separate from the negative prompt field where a model exposes one. On flow-matching architectures running at low guidance there is often no negative channel to push against at all, so a negative prompt of camera movement, shake, drift may be a straightforward no-op. The positive line in your main prompt is what you actually control.

    Put the stillness first

    Camera language placed at the end of a prompt gets dropped more often than camera language at the front, and "the camera is motionless" is camera language like any other. Lead with it. A shot card that opens Locked-off medium shot on a tripod, 50mm, the camera is entirely motionless. has set the premise before the model has decided anything about the scene. The same sentence appended after four clauses of subject and lighting description is asking the model to undo a camera it has already committed to.

    Staging tricks that help the frame stay put

    Explicit phrasing gets you most of the way. The rest is staging — describing a scene that has no reason to want a camera move.

    Give the movement to the subject. A locked frame with a static subject is a photograph, and models fill perceived stillness with camera drift. A locked frame with a subject who walks through it, or steam that rises, or a curtain that moves in the wind, has somewhere to put its motion budget. The camera is motionless. She walks from frame left to frame right and exits. is a far more stable request than The camera is motionless. She stands still.

    Name the mounting. "On a tripod," "clamped to the counter," "security camera mounted in the corner," "locked-off studio setup." These are not decorative. They describe a physical situation in which a camera move would be impossible, which is a stronger constraint than an adjective about the resulting image.

    Pick a frame type that implies stillness. Surveillance footage, a fixed webcam, a locked-off product plate on a seamless backdrop, a Wes-Anderson-style symmetrical tableau. Each of these is a genre where the still camera is the convention, and genre cues pull the whole generation toward the convention.

    Turn the motion dial down where one exists. Motion level is exposed as a real selectable value on part of the catalog and absent on the rest. Where it exists, low is the setting you want — but understand what it does. It reduces overall change per frame across subject, background and camera together. It is an amount dial, not a camera-versus-subject selector, so a low setting also flattens the subject motion you just staged. If your locked-off shot needs a subject who actually moves, leave the dial in the middle and carry the stillness in the prose.

    Keep the shot short. Drift accumulates. A three-second locked frame holds far more reliably than a ten-second one, and clip length is the single easiest variable to change. If you need ten seconds of a static frame, three clean clips that share framing will cut together more convincingly than one long generation that wanders in the back half.

    When prompting is not enough

    Some shots will not lock, no matter how you phrase them. Two ways out.

    Constrain both ends. First-and-last-frame generation lets you supply the start and end images yourself. Hand it the same framing at both ends and the model has nowhere to drift to, because the destination is pinned. This is the cleanest structural fix available for camera drift, and it works precisely because it stops being a prompting problem. Models including Veo's first-last-frame variant accept both endpoints directly.

    Fix it in the timeline. A shot that drifts slightly can often be cropped back to still, at the cost of some resolution. The video editor is EDL-based, so the timeline is re-renderable rather than baked — you can adjust and re-render without starting over. Previewing at 480p with preview: true is free and carries a short per-user cooldown, so you can check whether a crop actually holds before you spend anything on the final export, which is charged once regardless of how many clips are on the timeline.

    A quick test for whether your phrasing works

    Run three takes of your locked-off line with nothing else in the prompt but a simple subject and a plain background. Then watch only the frame edges, not the subject. Pick a fixed feature at the edge — a doorframe, a table corner, a horizon line — and see whether it stays at the same pixel across the clip. Subject motion in the middle of the frame will hide several pixels of camera drift from casual viewing; edge features will not.

    If two of three takes hold, your phrasing is working and you have a normal reroll rate to budget for. If none hold, the phrasing is not the problem and you should move to the first-last-frame route rather than rewriting the sentence a sixth time. The general version of this diagnostic — deciding when to keep prompting versus when to change technique — is laid out in the diagnostic tree for generations that come back wrong.

    FAQ

    Is a locked-off shot easier or harder for a model than a camera move?

    Easier, once you ask for it explicitly. Stillness maps to "do less," which is the cheapest thing a model can satisfy, whereas any named move asks for sustained specific behavior over the whole clip. The catch is entirely in the asking — the reliability advantage only shows up after you state the stillness, because silence gets filled with the model's default, which moves.

    Will a negative prompt of "camera shake, camera movement" fix drift?

    Not dependably, and on some architectures not at all. Where a model runs without a usable negative-guidance channel, that field is effectively ignored. Even where it works, instructive negation is weaker than positive re-specification. Write the motionless tripod into the main prompt instead.

    Why does my frame hold for four seconds and then start drifting?

    Accumulated error over the length of the generation. The model has less context anchoring the tail of a clip than the head, and small per-frame inconsistencies compound. Shorten the clip, or pin the end state with a first-and-last-frame generation so the tail has a target to hit.

    Does "static shot" work as well as spelling out that the camera does not move?

    Sometimes, and that inconsistency is the reason not to rely on it. "Static shot" is also read by some models as a description of the subject rather than the camera, which is how you end up with a frozen person and a wandering lens. Naming the camera as the thing that is motionless removes the ambiguity for one extra clause.