Home›AI Camera Movements

AI Camera Movements

Eighteen movements, each demonstrated on the same scene so the only variable is the camera. Every clip was generated with the prompt printed on its card.

Translation

Dolly In

A dolly in moves the entire camera forward toward the subject. Because the body travels through space, objects at different distances shift across each other and the shot gains genuine depth. That depth separation is the whole point, and it is what a zoom in cannot produce no matter how closely the two are matched.

slow dolly in toward the subject, camera pushes forward through the space, shallow depth of field

Full breakdown

Camera behaviour: The camera body translates forward along the axis the lens is pointing. Focal length is not touched, so the lens keeps the same angle of view while the distance to the subject closes.
Perspective and parallax: Strong. Foreground elements sweep past the frame edges faster than background elements, which is the visual signature of a camera that is genuinely moving through space.
Visual effect: Involvement. The viewer is drawn into the scene rather than shown a larger picture of it, which is why the move reads as attention narrowing rather than as a technical operation.
When to use it: Arriving at a realisation, entering a character's interior state, or committing to a subject the scene has been circling. Standard for the moment a scene stops surveying and starts focusing.
Prompting AI video: Use the words "dolly in" or "push in", and place them at the end of the prompt. Adding "shallow depth of field" reinforces the depth cue the model is being asked to simulate.
Common AI failure: Asking for "zoom in" instead. Models read zoom as a scale-up of the frame and return a flat digital enlargement with no depth separation — the single most common camera-prompt mistake.
How it differs: Versus a zoom in: watch the background. In a dolly the background shifts relative to the subject; in a zoom everything scales together and nothing shifts.
How it is done for real: Track and dolly, or a gimbal on a slider. On location a stabilised gimbal walked forward substitutes for track, at the cost of a slight vertical float.
Combines well with: Follows a static shot well: holding still first makes the push read as a decision rather than restlessness.

Translation

Dolly Out

A dolly out moves the entire camera backward, away from the subject, so surrounding space arrives progressively at the frame edges. Because the camera genuinely retreats, the new space enters with correct perspective rather than simply appearing, which is what separates it from a zoom out covering the same range.

slow dolly out away from the subject, camera retreats, revealing the surrounding environment

Full breakdown

Camera behaviour: The camera body translates backward along the lens axis. The angle of view is unchanged; the distance to the subject opens up.
Perspective and parallax: Strong, and in reverse. Foreground elements fall away toward the edges faster than the background, so the scene appears to reassemble itself around the subject.
Visual effect: Release, isolation or revelation, depending on what the widening exposes. The subject shrinks inside a world that turns out to be larger than the frame admitted.
When to use it: The closing move of a scene, and the standard way to reveal that a character is alone, surrounded, or smaller than they seemed. Also the classic final shot of a film.
Prompting AI video: Name what should be revealed. "Dolly out to reveal the empty field" gives the model a target for the widening; a bare "dolly out" gives it nothing to aim at.
Common AI failure: Given no reveal target, models frequently hold camera position and simply widen the frame — a zoom out wearing the wrong name, with the depth cue missing.
How it differs: Versus a zoom out: in a dolly out the foreground rushes away from the edges. In a zoom out the foreground stays exactly where it is, only smaller.
How it is done for real: The same track run in reverse, or a gimbal walked backward with a spotter. Crane arms are used when the retreat also needs to rise.
Combines well with: Pairs with a crane up to turn a personal moment into a landscape in one continuous move.

Translation

Tracking Shot

A tracking shot travels alongside a moving subject, holding roughly constant distance so the subject stays framed while the world streams past behind it. The camera and the subject move together as a pair, which is what keeps a composition stable through motion that would otherwise carry the subject out of frame.

tracking shot following the subject from behind at walking pace, camera moves with them, steady

Full breakdown

Camera behaviour: The camera body translates along the subject's path at a matched speed, from behind, alongside or in front. Distance and framing are held; the route is dictated by the subject.
Perspective and parallax: Strong, and it all lands in the background. The subject is nearly static in frame while everything behind it sweeps past — the clearest demonstration of parallax in the whole set.
Visual effect: Accompaniment. The viewer travels with the subject rather than watching them pass, which builds continuity and momentum through a scene.
When to use it: Walk-and-talks, chases, entrances, and any sequence where the geography between two points matters as much as the destination.
Prompting AI video: State the position and the pace: "following from behind at walking pace". Both are needed — position stops the model drifting into an orbit, pace stops it accelerating.
Common AI failure: Drift between three moves. Without a stated position, models slide between following, orbiting and simply panning as the subject leaves frame, often within a single clip.
How it differs: Versus a dolly: a dolly follows a path the camera chooses, usually toward or away. A tracking shot follows a path the subject chooses, keeping pace.
How it is done for real: Track, a gimbal, a vehicle mount, or a Steadicam when the route has to include stairs and corners.
Combines well with: Combines naturally with a whip pan at the end, handing the shot off to whatever the subject arrives at.

Translation

Truck Left

also called dolly left

A truck left slides the entire camera sideways to the left while the lens keeps pointing the same direction. Because the body travels, objects at different depths slide across one another. That parallax is the difference from a pan left, which sweeps the same direction from a fixed point and produces none.

camera trucks left, sliding laterally on a straight path, lens facing forward, smooth constant speed

Full breakdown

Camera behaviour: The camera body translates horizontally, perpendicular to the lens axis. The direction of view never changes — only the camera's position in space.
Perspective and parallax: Strong and unmistakable. Anything close to the lens crosses the frame far faster than the horizon, and a foreground object passing through sells the depth instantly.
Visual effect: Surveying. The scene is presented as a space with layers rather than as a picture, and the sideways drift reads as deliberate, composed observation.
When to use it: Revealing a room or a line-up, moving between two subjects at the same distance, and any shot where foreground objects should pass through to establish depth.
Prompting AI video: Use "trucks left" or "slides laterally to the left" and add "lens facing forward" so the model does not convert the move into a rotation.
Common AI failure: The model rotates instead of sliding, returning a pan. Adding "without turning the camera" corrects it more reliably than repeating the word truck.
How it differs: Versus a pan left: a truck moves the camera, a pan turns it. If the foreground and background stay locked together, you were given a pan.
How it is done for real: A slider for short moves, dolly track for long ones. The slider is the tool; the truck is the move — they are not two different shots.
Combines well with: Works with a rack focus: as the lateral move brings a new plane forward, the focus follows it.

Translation

Truck Right

also called dolly right

A truck right slides the entire camera sideways to the right with the lens still facing forward. The lateral travel separates foreground from background, which is the visible difference from a pan right: the same apparent direction of sweep, produced by completely different geometry and reading completely differently.

camera trucks right, sliding laterally on a straight path, lens facing forward, smooth constant speed

Full breakdown

Camera behaviour: Mirror of truck left. Horizontal translation perpendicular to the lens axis, view direction fixed throughout.
Perspective and parallax: Strong. Depth layers separate at different rates, and the effect is proportional to how close the nearest object is to the lens.
Visual effect: The same composed observation as truck left, with the opposite reading direction — which matters, because sweeping with the eye's reading direction feels settled and against it feels unresolved.
When to use it: Lateral reveals, moving along a wall, shelf or crowd, and matching a subject's movement without following them.
Prompting AI video: Identical to truck left with the direction reversed. Keep "lens facing forward" — it is doing the real work in the prompt.
Common AI failure: Direction reversal mid-clip. Where the scene has no strong directional cue, models sometimes settle the move by drifting back the other way.
How it differs: Versus a pan right: identical sweep direction, opposite mechanics. Only the truck moves the camera through space.
How it is done for real: Identical hardware to truck left. Direction is a staging decision, not a technical one.
Combines well with: Pairs with a static subject so the parallax has something fixed to be measured against.

Rotation

Pan Left

A pan left rotates the camera horizontally to the left from a fixed position. Only the angle changes, so nothing moves relative to anything else and no parallax appears. This is precisely the difference from a truck left, where the body travels sideways and the depth layers visibly separate.

slow pan left, camera rotates horizontally from a fixed position, level horizon

Full breakdown

Camera behaviour: Rotation about the vertical axis running through the camera. Position is unchanged; the frame sweeps across the scene like a turning head.
Perspective and parallax: None. Because the camera does not change position, the spatial relationships in the scene are frozen and the image reads flat in depth.
Visual effect: Surveying without travelling. The viewer is held in one place and shown what surrounds it, which reads as observation rather than participation.
When to use it: Establishing geography from a fixed vantage, following action across a static frame, or connecting two subjects that share a space.
Prompting AI video: Add "from a fixed position" explicitly. It is the single most useful phrase in the prompt, because the default failure is translation.
Common AI failure: The model translates instead of rotating, quietly converting the pan into a truck. Naming an end point ("pan left to settle on the doorway") also stops the rotation drifting past its target.
How it differs: Versus a truck left: the pan produces no parallax at all. If the foreground separates from the background, the camera moved and you did not get a pan.
How it is done for real: A fluid head on a tripod. The whole point is that the base never moves, which is why a locked-off tripod is the tool.
Combines well with: Follows a tilt to survey a space in two axes without ever moving the camera position.

Rotation

Pan Right

A pan right rotates the camera horizontally to the right from one spot, bringing new space in at the right edge. Because the camera never travels, the image stays flat in depth — the reliable tell that separates a pan from a lateral truck move sweeping the same direction.

slow pan right, camera rotates horizontally from a fixed position, level horizon

Full breakdown

Camera behaviour: Mirror of pan left. Rotation about the vertical axis, camera position fixed.
Perspective and parallax: None. Angle changes; geometry does not.
Visual effect: A settled, natural sweep — it follows the reading direction, so it feels like the scene resolving rather than opening a question.
When to use it: Following a subject across a fixed frame, revealing a second subject, or closing a scene by turning away from it.
Prompting AI video: Give the rotation a destination. Models handle "pan right to the treeline" far more consistently than an unbounded "pan right".
Common AI failure: Under-rotation. Given no target, models often turn a few degrees and stop, producing something closer to a static shot with drift.
How it differs: Versus a truck right: watch a near object. In a pan it holds its position relative to the background; in a truck it does not.
How it is done for real: A fluid head, damped so the rotation starts and ends without a jolt. Speed is controlled by drag, not by the operator's arm.
Combines well with: Combines with a subject entering frame, so the rotation appears motivated rather than arbitrary.

Rotation

Tilt Up

A tilt up rotates the camera upward from a fixed position. The body stays exactly where it is and only the angle rises. That is the whole difference from a pedestal up, which physically lifts the camera while the lens keeps pointing straight ahead — same upward reading, entirely different geometry.

slow tilt up, camera pivots vertically from a fixed position, revealing the height of the subject

Full breakdown

Camera behaviour: Rotation about the horizontal axis through the camera. Height is constant; the frame swings upward, and vertical lines converge more strongly as the angle steepens.
Perspective and parallax: None. The camera holds its position, so the scene cannot separate into depth layers — only the angle onto it changes.
Visual effect: Scale and elevation. Revealing height from below makes the subject monumental, because the converging verticals are themselves a cue of looking up.
When to use it: Revealing the height of architecture or a figure, moving from detail to sky, or opening a scene by lifting the eye from the ground.
Prompting AI video: State where the tilt lands. "Tilt up to the roofline" is dramatically more reliable than "tilt up" alone.
Common AI failure: The tilt stops a few degrees in. Models under-commit to vertical rotation more than to any other move, so the destination phrase is doing real work.
How it differs: Versus a pedestal up: the tilt changes the angle, the pedestal changes the height. In a tilt the perspective distortion increases; in a pedestal it does not.
How it is done for real: A fluid head with counterbalance. Without it the head falls under the lens's own weight and the tilt accelerates at the end.
Combines well with: Pairs with a dolly in for a rising push — but ask AI models for them one at a time.

Rotation

Tilt Down

A tilt down rotates the camera downward from a fixed position, swinging the frame toward the ground without lowering the camera itself. Height never changes — only the angle — which is exactly what distinguishes it from a crane or pedestal descending through the same arc.

slow tilt down, camera pivots vertically from a fixed position, settling on the subject below

Full breakdown

Camera behaviour: Mirror of tilt up. Rotation about the horizontal axis, camera height fixed.
Perspective and parallax: None. Position is unchanged, so no depth separation is possible.
Visual effect: Diminishment or arrival. Looking down at a subject reduces it; looking down from height to a detail concentrates attention sharply.
When to use it: Landing on a subject after establishing context above it, moving from sky to ground, or ending a sequence by dropping the eye.
Prompting AI video: Combine with nothing else at first. A tilt paired with a push is two instructions, and models routinely perform only the easier one.
Common AI failure: Silent substitution with a crane down. If the horizon rises through the frame while the perspective stays flat, the model moved the camera instead of turning it.
How it differs: Versus a crane down: the tilt only changes the angle. If the camera's height above the ground visibly changes, it was a crane.
How it is done for real: The same counterbalanced head, run the other way. On a jib the tilt is often automated to hold the subject as the arm moves.
Combines well with: Follows a crane up, so the descent of the eye answers the rise of the camera.

Compound

Orbit

also called an arc shot

An orbit travels the camera around the subject on a curved path while continuously turning to keep it framed. The subject stays roughly centred as the background rotates behind it. That moving background is the signature, and a pan from a fixed point can never reproduce it.

camera orbits around the subject, smooth continuous arc, subject stays centered in frame

Full breakdown

Camera behaviour: Translation along a circular path combined with continuous rotation inward, so the lens keeps facing the centre while the camera position changes constantly.
Perspective and parallax: Continuous and total. The entire background is replaced over the course of the move while the subject holds still, which is the most complete depth demonstration available.
Visual effect: Examination and importance. Circling a subject presents it as an object with volume, and the accumulating background change builds a sense of significance.
When to use it: Hero shots, product reveals, and the moment a scene stops observing a character and starts studying them.
Prompting AI video: Give a direction and an amount: "orbit clockwise, roughly 180 degrees". Both are needed — direction prevents reversal, amount prevents the move from stalling.
Common AI failure: Reversal mid-clip, and collapse into a pan. An unbounded orbit frequently changes direction halfway, or degrades into rotation on the spot with a static background.
How it differs: Versus a pan: in an orbit the background changes completely while the subject holds. In a pan the subject crosses the frame and the background stays put.
How it is done for real: A circular track, a gimbal walked around the subject, or a drone in an automated point-of-interest mode.
Combines well with: Pairs with a slow push, tightening the arc into a spiral as it goes round.

Compound

Crane Up

also called jib up, boom up

A crane up lifts the camera through space on an arm, usually rising and drifting forward at the same time. Scale is what marks it: the viewpoint climbs well beyond human height, converting a ground-level subject into one element of a much wider landscape.

crane shot rising from ground level up and over the scene, sweeping vertical reveal

Full breakdown

Camera behaviour: Vertical translation on an arm, normally combined with forward travel and a compensating downward tilt so the subject stays framed as the camera climbs.
Perspective and parallax: Strong and vertical. The ground plane opens out below while the horizon stays put, and foreground objects drop away rapidly at the bottom of the frame.
Visual effect: Transcendence and conclusion. Rising above a scene ends involvement in it, which is why the move so often closes a sequence.
When to use it: Establishing shots that begin intimate, the final shot of a scene, and any reveal where the answer is "there is far more here than you could see".
Prompting AI video: "Crane up and over" produces a much stronger arc than "crane up", which models tend to flatten into a straight vertical rise.
Common AI failure: Flattening into a pedestal. Without the forward component the model delivers a straight lift, losing the sweep that makes a crane read as a crane.
How it differs: Versus a tilt up: the crane changes the camera's actual height. A tilt only raises the angle, and the ground plane never opens out.
How it is done for real: A jib or crane arm, or a drone taking off from a low hover. The arm gives a bounded, repeatable arc that a drone does not.
Combines well with: Ends a dolly out well: the retreat becomes an ascent and the scene closes itself.

Compound

Crane Down

also called jib down, boom down

A crane down descends through space on an arm, typically dropping from a high vantage toward a subject at ground level. Height and framing change together, which is why it reads as an arrival into a scene rather than the simple change of angle a tilt down produces.

crane shot descending from above down toward the subject, smooth sweeping fall

Full breakdown

Camera behaviour: Vertical downward translation on an arm, usually with forward travel and a compensating upward tilt as the camera settles.
Perspective and parallax: Strong. The ground plane closes up as the camera descends and foreground elements rise into frame from below.
Visual effect: Arrival and entry. Descending into a scene commits the viewer to it, the exact inverse of the crane up's withdrawal.
When to use it: Opening a scene from an establishing height, landing on a character after context, or handing off from a wide to a conversation.
Prompting AI video: Name the landing point. A crane down with no destination tends to stop halfway and hover at an arbitrary height.
Common AI failure: Stalling. Models decelerate a descent early far more often than they overshoot it, leaving the move visibly unfinished.
How it differs: Versus a tilt down: the crane physically loses altitude. If the camera's height is constant and only the angle drops, it was a tilt.
How it is done for real: The same arm, descending, usually with the head tilting up to keep the subject framed as the camera drops.
Combines well with: Leads into a tracking shot, arriving at a subject and then travelling with them.

Optical

Zoom In

A zoom in increases focal length while the camera stays completely still. Nothing travels through space, so no parallax appears and the image simply magnifies. Perspective is untouched, which is why a zoom reads flat next to a dolly in that covers the same apparent distance.

slow zoom in on the subject, focal length increases, camera position stays fixed

Full breakdown

Camera behaviour: A lens-only change. The angle of view narrows; the camera's position, and therefore the perspective, are not altered at all.
Perspective and parallax: None, by definition. Every part of the frame scales at exactly the same rate, and the compressed background is the visible consequence.
Visual effect: Pointing rather than approaching. The move directs attention without moving the viewer, which can feel clinical, sudden or observational depending on speed.
When to use it: Deliberate emphasis, documentary and surveillance registers, and the compressed telephoto look where background separation is wanted without travel.
Prompting AI video: This is the one case where "zoom" is the correct word. Say "focal length increases, camera position fixed" to stop the model adding travel it was not asked for.
Common AI failure: Being used when a dolly was wanted. The output looks flat and digital, and no amount of prompt refinement fixes it, because the flatness is what a zoom is.
How it differs: Versus a dolly in: identical framing change, opposite physics. The zoom compresses the background; the dolly moves through it.
How it is done for real: A zoom lens, cranked by hand or by a servo. Nothing else in the rig moves, which is exactly why the shot reads as it does.
Combines well with: Combines with a dolly out to produce a dolly zoom — the only case where a zoom belongs with a movement.

Optical

Zoom Out

A zoom out decreases focal length from a fixed camera position, widening the frame without moving. More of the scene becomes visible, but every spatial relationship stays exactly as it was — the flatness that distinguishes it from a dolly out, where the camera genuinely retreats through the space.

slow zoom out from the subject, focal length decreases, camera position stays fixed

Full breakdown

Camera behaviour: A lens-only change in the opposite direction. The angle of view widens; camera position and perspective are unchanged.
Perspective and parallax: None. The frame scales uniformly; nothing separates from anything else.
Visual effect: Context without withdrawal. The viewer is shown more while staying exactly where they were, which reads as information rather than as movement.
When to use it: Adding context quickly, comic or shock reveals where the speed of the widening is the point, and correcting framing without a rig move.
Prompting AI video: More reliable than zoom in, because widening tolerates invented edge detail better than magnification tolerates invented centre detail.
Common AI failure: Edge invention. As the frame widens the model must produce scene that was never established, and inconsistency shows up at the borders first.
How it differs: Versus a dolly out: in a zoom out the foreground stays put and shrinks. In a dolly out it rushes away toward the edges.
How it is done for real: A zoom lens run the other way. Servo control matters more here, because uneven speed is obvious as the frame opens.
Combines well with: Pairs with a dolly in for the reverse dolly zoom, where the background rushes toward the subject.

Platform

360 Roll

also called a barrel roll or Dutch roll

A 360 roll rotates the camera around the lens axis, turning the horizon through a full revolution. It is the third rotation axis — neither pan nor tilt — and it is the only movement that makes the frame itself feel unmoored rather than the subject inside it.

camera rolls 360 degrees around the lens axis, horizon rotates fully, subject stays centered

Full breakdown

Camera behaviour: Rotation about the axis the lens points along. Position and framing are unchanged; only the horizon angle turns.
Perspective and parallax: None. The camera does not change position, so the scene rotates as a single rigid image.
Visual effect: Disorientation, exhilaration or unreality, depending on speed. Because the horizon is a fixed reference for human balance, turning it is felt physically.
When to use it: Transitions, dream and altered states, action beats, and music video work where the frame is treated as an object in its own right.
Prompting AI video: Say "roll", never "rotate" — most models read "rotate" as an orbit. Adding "horizon rotates" removes the remaining ambiguity about which axis is meant.
Common AI failure: Delivered as an orbit. This is the most reliably misread instruction in the set, because "rotate" maps to the more common movement in training data.
How it differs: Versus an orbit: the roll turns the image and keeps the background fixed. The orbit keeps the image upright and replaces the background.
How it is done for real: A gimbal with an unlocked roll axis, a drone, or a full camera rotation rig. Handheld cannot produce a clean full revolution.
Combines well with: Best against a static subject: the roll is only legible if something in frame stays put.

Platform

Handheld

A handheld shot is carried by an operator rather than mounted on a rig, so small corrections and instabilities stay in the image instead of being engineered out. The motion is not decoration: that unsteadiness reads as presence, which is why documentary and confrontation scenes reach for it deliberately.

handheld camera, natural micro-movement and slight shake, documentary feel, following the subject

Full breakdown

Camera behaviour: No fixed platform. Small translations and rotations combine continuously and irregularly, with the operator's body as the suspension.
Perspective and parallax: Present but small and constant. The micro-translations produce a continuous low-level depth shimmer, which is a large part of why handheld reads as real.
Visual effect: Immediacy and unease. The frame feels observed rather than composed, and the viewer reads the instability as a human being present.
When to use it: Documentary, conflict, urgency, and any scene that should feel witnessed rather than staged.
Prompting AI video: Ask for "subtle" or "natural" movement explicitly. The qualifier matters more than the word handheld.
Common AI failure: Exaggeration. Unqualified "handheld" pushes many models into a wobble that reads as an applied effect rather than a person holding a camera.
How it differs: Versus a static shot: handheld never fully settles. Versus a tracking shot: handheld describes the suspension, not the path.
How it is done for real: The operator's body, sometimes with a shoulder rig or an easyrig to take the weight without removing the movement.
Combines well with: Pairs with a tracking intent — following a subject handheld is the documentary default for good reason.

Platform

Aerial

also called a drone shot

An aerial shot moves the camera through the air above the scene, translating in three dimensions at a scale no ground rig can reach. Altitude is the point: it converts individual subjects into pattern and geography, and it establishes a place faster than any amount of ground-level coverage.

aerial drone shot flying over the landscape, high altitude, smooth forward movement

Full breakdown

Camera behaviour: Free translation on all three axes, usually with a fixed downward or forward camera angle held by a gimbal.
Perspective and parallax: Strong, and unusually legible because the ground plane is the reference. Everything below shifts against a fixed horizon as the camera travels.
Visual effect: Scope and detachment. The viewer is placed outside the human vantage entirely, which reads as omniscience rather than participation.
When to use it: Establishing shots, landscape and architecture, and any moment where the relationship between subject and geography is the story.
Prompting AI video: Specify altitude and direction of travel. "Aerial" alone frequently returns a static high-angle frame with no movement at all.
Common AI failure: No travel. The model produces the vantage without the motion, which is a high-angle static shot rather than an aerial.
How it differs: Versus a crane up: the crane is anchored and its arc is bounded. An aerial is free in all three axes and unbounded in range.
How it is done for real: A drone. Larger productions use a helicopter with a gyro-stabilised mount, which buys altitude and endurance, not steadiness.
Combines well with: Combines with a tilt down as the aircraft passes over, turning a fly-over into a reveal.

Platform

Static

also called a locked-off shot

A static shot does not move at all. It belongs in any list of camera movements because it is the baseline the other seventeen are read against: when the frame holds still, every movement inside it becomes the subject, and holding still is frequently the stronger choice.

static locked-off shot, camera does not move at all, tripod, subject moves within the frame

Full breakdown

Camera behaviour: No translation, no rotation, no change of focal length. The only motion in the shot belongs to the scene.
Perspective and parallax: None, and its absence is the point. With the camera fixed, depth must be communicated by composition, lens choice and staging instead.
Visual effect: Attention and stillness. A held frame lets the viewer look rather than be led, and it makes any later movement land much harder.
When to use it: Dialogue, observational comedy, tension built through waiting, and as the contrast that gives a moving shot its impact.
Prompting AI video: Say "locked off" and "camera does not move". Stillness is an instruction, not the absence of one — video models drift by default.
Common AI failure: Drift. Almost every model adds slow movement unless told not to, because motion is what the training objective rewards.
How it differs: Versus handheld: both can appear to hold framing, but a static shot has zero movement while handheld never stops correcting.
How it is done for real: A tripod with the head locked, or any rigid mount. Also the hardest to achieve in AI video, where drift is the default.
Combines well with: Pairs with everything, by contrast: a held frame is what makes the next moving shot land.

Grouped by what actually moves

A camera can travel through space, pivot on the spot, or stay put while the lens changes. That single fact decides how a shot reads, and nearly every confusion in this subject traces back to it. When the body travels, objects at different distances shift across each other — parallax — and the image gains depth. When it only pivots, no such shift occurs and the image stays flat.

Translation

The camera body travels through space. Parallax appears.

Dolly In
Dolly Out
Tracking Shot
Truck Left
Truck Right

Rotation

The body pivots on the spot. No parallax.

Pan Left
Pan Right
Tilt Up
Tilt Down

Compound

Travel and rotation at once.

Orbit
Crane Up
Crane Down

Optical

The camera holds still; the lens changes.

Zoom In
Zoom Out

Platform

Defined by what carries the camera.

360 Roll
Handheld
Aerial
Static

The five comparisons that matter

These five pairs account for most of the confusion around camera movement, and they are where published guides most often go wrong. Each one comes down to the same question: does the camera travel through space, or does it stay where it is?

What is the difference between a dolly and a zoom?

A dolly moves the camera through space; a zoom changes the lens from a fixed position. Both make the subject larger, and there the similarity ends. The dolly produces parallax and reveals depth. The zoom magnifies uniformly and compresses the background, so the image reads flat.

Side by side

Dolly In

The camera body travels forward. Foreground and background separate, perspective changes, and the viewer feels moved into the scene.

Zoom In

The camera never moves. Focal length narrows the angle of view, everything scales together, and the background compresses toward the subject.

How to tell them apart Watch the background against the subject. If their relationship changes, it is a dolly. If the whole frame scales as one image, it is a zoom.

In AI video This distinction costs nothing to get right and is the most valuable single fix in AI video prompting: write "dolly in" when you want depth, "zoom in" only when you want flat magnification.

Is a pan the same as a truck?

No. A pan rotates the camera from one fixed spot. A truck slides the whole camera sideways through space. Both sweep the frame horizontally, but only the truck produces parallax, and in motion the two are not remotely similar.

Side by side

Pan

Rotation on the vertical axis. Position is fixed, depth relationships are frozen, and the frame turns like a head.

Truck

Lateral translation. The lens keeps facing forward while the body travels, so near objects cross the frame faster than far ones.

How to tell them apart Pick something close to the lens. In a pan it keeps its position against the background. In a truck it slides across it.

In AI video Models default to translating when asked to pan. "From a fixed position" is the phrase that holds the camera still.

What is the difference between a tilt and a pedestal?

A tilt rotates the camera up or down from a fixed height. A pedestal raises or lowers the camera body itself while the lens keeps pointing the same way. The frame moves vertically in both cases, but only the pedestal changes where the camera actually is.

Side by side

Tilt

Angle changes, height does not. Vertical lines converge more strongly as the tilt steepens — the visual cue of looking up or down at something.

Pedestal no clip yet

Height changes, angle does not. Vertical lines stay parallel because the camera keeps looking straight ahead; the viewpoint simply rises or falls.

How to tell them apart Look at the vertical lines. If they lean and converge, the camera tilted. If they stay upright while the view rises, the camera was raised.

In AI video Pedestal is the weaker term in most models' vocabulary. "Camera rises vertically, lens stays level" describes the move more reliably than the word itself.

What is the difference between an orbit and a pan?

An orbit travels around the subject on a curved path, so the background is progressively replaced while the subject stays centred. A pan rotates from a fixed point, so the subject crosses the frame while the background stays exactly where it is. They are close to opposites.

Side by side

Orbit

The camera moves; the subject holds still in frame. Everything behind it rotates continuously, revealing the subject as an object with volume.

Pan

The camera turns; the subject moves across frame. The background is fixed and merely enters and leaves at the edges.

How to tell them apart Ask what stays still. Orbit holds the subject and moves the world. Pan holds the world and moves the subject through frame.

In AI video An unbounded orbit often collapses into a pan, because rotating on the spot is the cheaper solution. Naming an arc size ("roughly 180 degrees") prevents it.

What is the difference between a tracking shot and a dolly shot?

Both move the camera through space, so both produce parallax. The difference is who decides the path. A dolly follows a route the camera chooses, usually straight toward or away from the subject. A tracking shot follows the subject, matching their speed and direction.

Side by side

Tracking

Subject-led. The camera keeps pace with something already moving, holding distance and framing while the world streams past behind.

Dolly

Camera-led. The move exists independently of whether the subject moves at all, and is usually along the lens axis.

How to tell them apart Ask whether the shot would still make sense with a motionless subject. A dolly would. A tracking shot would have nothing to track.

In AI video They overlap in practice, and a tracking shot is often executed on a dolly. Prompt for the intent — "following the subject" versus "pushing in" — not for the equipment.

How to prompt camera movement

Four rules that hold across models. If your model exposes camera movement as a structured parameter rather than reading it from prompt text, use that instead — it is more reliable than any phrasing.

Put the camera instruction last

Models build the scene first and apply the camera afterwards. A movement named at the start of a prompt is usually absorbed by the scene description before it is ever applied.

Two or three modifiers, no more

Beyond that they compete, and the model resolves the conflict by performing the simplest one and discarding the rest.

Name the destination

"Tilt up" often produces a few degrees of drift. "Tilt up to the roofline" gives the movement somewhere to end, and the result is far more consistent.

Say what should not change

"From a fixed position" for a pan, "lens facing forward" for a truck, "camera does not move" for a static shot. Video models drift by default, so stillness is an instruction.

About these clips

Every clip on this page was generated by Leopati with an AI video model, using the prompt printed on its card. None of it is licensed stock footage standing in for a movement it might not perform.

They also all use the same scene: the same subject, the same field, the same light. That is deliberate. When each movement is demonstrated on different footage, the footage becomes a variable and the comparison stops being a comparison. Here the only thing that changes from one card to the next is the camera.

What are the main types of camera movement?

They divide by what physically moves. The camera body can travel through space (dolly, truck, tracking, crane, aerial), it can pivot in place without travelling (pan, tilt, roll), or it can stay completely still while the lens changes (zoom, focus pull). Everything else is a combination of those, or a description of what is carrying the camera.

What is the difference between a dolly and a zoom?

A dolly moves the camera through space, so objects at different distances shift at different rates and the shot gains depth. A zoom changes focal length from a fixed position, so nothing shifts relative to anything else and the image reads flat. They look similar in a still frame and completely different in motion.

Is a pan the same as a truck?

No. A pan rotates the camera from one fixed spot. A truck slides the whole camera sideways through space. Both sweep the frame horizontally, but only the truck produces parallax between foreground and background, and that is visible immediately in motion.

What is the difference between a tilt and a pedestal?

A tilt rotates the camera up or down from a fixed height. A pedestal raises or lowers the camera itself while the lens keeps pointing the same way. Look at the vertical lines: if they lean and converge, the camera tilted; if they stay upright while the view rises, the camera was raised.

What is the difference between an orbit and a pan?

An orbit travels around the subject, so the subject stays centred while the background is progressively replaced. A pan rotates from a fixed point, so the subject crosses the frame while the background stays put. Ask what stays still: the orbit holds the subject, the pan holds the world.

What is the difference between a tracking shot and a dolly shot?

Both move the camera through space. The difference is who decides the path. A dolly follows a route the camera chooses, usually toward or away from the subject. A tracking shot follows the subject, matching their speed and direction. A dolly still makes sense with a motionless subject; a tracking shot has nothing to track.

Why does my AI video look flat when I ask for a zoom?

Because that is what a zoom is. Zooming magnifies from a fixed position and cannot create depth. If you wanted the shot to feel like it is moving into the scene, ask for a dolly in or a push in instead, and the model will simulate the parallax that makes the movement read as travel.

How many camera instructions should one prompt contain?

Two or three at most, placed at the end of the prompt. More than that and the instructions compete; models typically resolve the conflict by performing the simplest one and ignoring the others.

Were these example videos filmed or generated?

Generated. Each clip was produced by Leopati with an AI video model, using the prompt shown on the same card. They are demonstrations of what these prompts actually return, not licensed stock footage selected to resemble a movement.