The single skill that separates a usable AI-generated shot from drifting slop is camera control — telling the model how to frame and move, not just what to show. In every major text-to-video model, that control is a language: a prompt written in real cinematography terms — shot size, camera angle, a specific movement, a lens — produces a directed shot, while a prompt that describes only the subject gets you whatever default the model reaches for, usually a static frame or an aimless drift. This guide is a working reference for that language. It covers the four layers of the vocabulary models actually respond to (framing, angle, movement, lens), the motion-prompting rules that make a move land reliably — one move per shot, positive phrasing, matching the move to the subject — and the one tool that inverts the whole approach by letting you pick a named camera move before you describe anything. It reads the four most-used models honestly on how much camera control each one really gives you, shows how to chain shots so a sequence holds continuity, and ends on the part most tutorials skip: what happens to a beautifully-directed clip after it renders, and the engine that turns it into finished, on-brand, scheduled content across every platform instead of a one-off download.
Ask a text-to-video model for 'a woman walking through a market' and you get a clip that technically matches and feels like nothing — a drifting frame, a camera that wanders, a shot no cinematographer would have chosen. Ask for 'a low-angle tracking shot following a woman through a crowded market, shallow depth of field' and you get a directed image. The subject barely changed. What changed is that the second prompt told the camera what to do. That gap is the entire skill of working with AI video in 2026: the models are good enough that the bottleneck is no longer whether they can render a scene, it is whether you can direct the camera through it.
The reason this matters more than it used to is that the subject is now the easy part. Modern models render coherent people, lighting, and physics on the first try far more often than the early tools did, so the difference between amateur and professional output has moved to the camera. A shot reads as 'AI slop' not because the pixels are wrong but because nobody decided where the camera was, how big the subject should be in frame, or how the shot should move — so the model decided, badly. Learning the camera vocabulary is how you take that decision back. The companion pieces on AI video generation in 2026 and AI video production cover the broader workflow; this guide is specifically about the language of the shot.
Before the vocabulary, understand that there are two fundamentally different control surfaces, and which one you are using changes how you work. The dominant approach is prompt-based: you write the camera direction into the text prompt in cinematography terms and the model interprets it. Runway, Google's Veo, Kling, and OpenAI's Sora all work primarily this way. It is maximally flexible — any shot you can describe, you can attempt — but it is also probabilistic, because the model is interpreting language, so the same prompt can produce slightly different moves across runs.
The second surface is preset-based, and it inverts the order of operations. Instead of describing a move, you select a named camera move from a menu before you write anything about content. This is Higgsfield's core design, and it trades the open-endedness of prompting for repeatability: a 'Crash Zoom' preset performs the same camera path every time, so the move is a known quantity and the prompt only has to carry the scene and the timing. Neither surface is strictly better. Prompting wins for arbitrary, bespoke shots; presets win when you want a specific cinematic move to be reliable and consistent across many clips. Most of this guide is about the prompt-based vocabulary, because that is what transfers across every model — but the preset approach gets its own section because for a certain class of shot it is simply more dependable.
Camera control in a prompt is not free-form description — it is a stack of four specific layers, and models respond best when you touch each one deliberately. Think of it the way a cinematographer briefs a shot: how big is the subject in frame, where is the camera relative to it, how is it moving, and what glass is it shot on. Naming all four is what turns a vague request into a directed shot. You do not need to specify every layer on every shot, but the ones you leave unspecified are the ones the model will decide for you.
Shot size is how much of the subject fills the frame, and it is the single most impactful control because it sets the emotional distance of the shot. The standard ladder runs from the extreme close-up (an eye, a detail) through close-up, medium shot (roughly waist-up), wide shot (the full figure in its environment), to the extreme wide or establishing shot (the subject small inside a large space). The filmmaking shorthand — ECU, CU, MS, WS — is widely recognized by the models, so 'medium close-up of a barista' is a valid and precise instruction. Choosing the size is a storytelling decision: push close for intimacy or tension, pull wide to establish place or scale.
Camera angle is where the lens sits relative to the subject's eyeline, and small changes carry large meaning. Eye-level reads as neutral and natural; a low angle (looking up at the subject) conveys power or dominance; a high angle (looking down) conveys vulnerability or smallness; an overhead or bird's-eye view flattens a scene into a graphic. Position terms layer on top — 'from the side,' 'over-the-shoulder,' 'point-of-view,' 'ground level' — and these are reliably understood. An over-the-shoulder two-shot or a POV frame tells the model not just how big the subject is but where the viewer is standing, which is often the difference between a shot that feels composed and one that feels floating.
Movement is where motion prompting lives, and it has the most terms because real cameras move in distinct, named ways. A pan is horizontal rotation on a fixed point (left or right); a tilt is vertical rotation (up or down). A dolly moves the whole camera toward or away from the subject — dolly-in (or push-in) to build focus and tension, dolly-out (pull-out) to reveal context. A tracking shot follows a moving subject to keep it in frame. An orbit (or arc) circles around a subject. A crane or pedestal moves the camera vertically through space. Handheld adds deliberate instability for an organic, documentary feel. And 'static' or 'locked' holds the frame completely still. These are not decorative labels — each produces a visibly different, recognizable move, and naming the right one is more than half of directing a shot.
The last layer is the optics, and it governs the look as much as the move. Focal length changes the feel of space: a wide lens (think 18–24mm) exaggerates depth and distance, a long lens (85mm and up) compresses it and isolates a subject. Depth of field controls what is sharp — 'shallow depth of field' or 'f/1.4' throws the background into a soft blur that reads as cinematic and directs the eye; deep focus keeps everything crisp. Specialty terms like anamorphic (for a widescreen, flared look) are recognized too. Lens language is how you make a shot look like it was photographed rather than rendered, and it is frequently the layer beginners skip — which is exactly why their output looks flat.
Knowing the vocabulary is necessary but not sufficient; there are a handful of prompting rules that decide whether the model honors your direction or averages it into drift. The first is the most important: one primary move per shot. Models interpret language and then try to satisfy all of it, so asking for a pan and a zoom and a tilt in one clip usually produces a muddy blend of all three. Pick the one move that defines the shot and commit to it. If you genuinely need a compound move, that is a sign to split it into two shots and cut between them, not to cram it into one prompt.
The second rule is specificity. 'The camera moves' gives the model almost nothing to work with and invites the default drift; 'a slow push-in' gives it a direction, a speed, and a type. Generic motion verbs are the single most common cause of a shot that ignores you. The third rule is positive phrasing, and it catches people constantly: most of these models do not reliably honor negations. 'No camera movement' or 'no shake' tends to be ignored or even inverted — the model latches onto 'movement' and 'shake.' Say what you want instead: 'the camera remains completely still,' 'locked tripod shot.' This single reframe fixes a surprising share of 'why won't it hold still' problems.
The fourth rule applies to image-to-video specifically: match the move to the image. If your starting frame is a calm portrait, a locked camera that lets subtle motion breathe will look right, while an aggressive orbit will fight the composition. Align the camera direction with what the image already implies. A useful fifth habit, reported to reduce failed generations, is to give a motion an endpoint — 'the camera pushes in, then settles' — so the model has somewhere to land the move rather than running it off the edge of the clip. None of these rules are model-specific folklore; they fall out of the same fact, that you are directing a language model's interpretation of a camera, and clear, singular, positive, grounded instructions interpret better than vague ones.
For a whole class of shot — the signature cinematic move you want to look identical every time — prompting is the wrong tool, and this is the gap preset-driven control fills. Higgsfield built its product around this inversion: rather than a text box you describe a move into, its primary surface is a gallery of named cinematographer moves — Dolly In, Crane Up, Head Tracking, Whip Pan, Bullet Time, Crash Zoom In, 360 Orbit, FPV Drone — that you pick first. The preset encodes the camera path, so you are no longer hoping the model interprets 'crash zoom' correctly; you are selecting a move that performs the same way on every generation. The prompt's job narrows to the content and the timing of the beat.
The trade-off is real and worth stating honestly. Presets give you repeatability and a guaranteed recognizable move, which is exactly what you want for a punchy hook or a consistent brand look, but they constrain you to the moves in the library and to short clips — this style of generation targets a few seconds of a single strong move, not a long, evolving take. The practitioner's discipline with these is restraint: a crash zoom or a bullet-time freeze is a one-beat-per-clip effect, and overusing it kills the impact. Higgsfield also layers a more manual production mode on top with explicit controls for lens, focal length, aperture, and move-set for people who want to leave the preset gallery — but the preset-first idea is the part worth stealing conceptually, because it clarifies that some camera moves should be chosen, not described.
The models differ less in vocabulary — they mostly share the same cinematography terms — than in how much control they give and how reliably they honor it. Runway's Gen-4.5 line prompts for movement style (locked, handheld, dolly, pan and more), can track a subject or move independently through an environment, and handles focus shifts; its guidance steers you toward filmic motion terms and one primary move at a time, and its physics simulation means the lighting shifts with the camera, so describing the light matters. Google's Veo (the 3.x line) is strong on explicit camera language, but prompting guides consistently report that it tends toward a static shot or a simple pan when a prompt gives no camera direction at all — with Veo, the safer assumption is that if you want the camera to move, you have to say so, clearly.
Kling's recent 3.0 models put camera movement more at the center and respond to the standard push/pull/pan/tilt vocabulary, with the community consensus that simple, explicit phrasing beats elaborate cinematic prose and that a scene described in a few concrete clauses outperforms a vague one-liner. OpenAI's Sora reads cinematic language from the prompt rather than exposing camera dials, so the same motion-prompting discipline applies. And Higgsfield, covered above, is the outlier that leads with presets. The practical takeaway is that the vocabulary in this guide transfers across all of them — what changes is the default behavior and the reliability, so the move is to learn the language once and then calibrate how explicit and how restrained you need to be per model. The deeper per-model breakdown, including where each sits on quality and cost, is in AI video generators for creators and AI video generation model tiers.
A single directed shot is a clip; a sequence of them is a film, and the hard part of a sequence is continuity. The newer models increasingly support multi-shot prompting — chaining several shots with different angles and framings while trying to hold the same subject and space — but it only works if you feed it continuity. The technique is to keep the subject and environment descriptions byte-for-byte identical across the shots and vary only the camera: the same 'a chef in a warm, steam-filled kitchen' described in a wide establishing shot, then a medium from the side, then a close-up on the hands, reads as one scene from three angles. Change the wording of the subject between shots and the model is liable to render three different chefs.
The classic structure is to open on a master or establishing shot — a wide, static or slow frame that shows the whole space and everyone in it — and then cut in to tighter coverage that reuses the exact same scene description. This mirrors how a real shoot is blocked, and it works for the same reason: the wide shot anchors the geography, and the coverage borrows from it. When continuity breaks anyway — a common failure — the fix is almost always to tighten the shared description and reduce how much changes between shots, not to add more camera instruction. Continuity is a discipline of repetition, not of elaboration.
Here is the honest line this whole guide builds to: camera control and motion prompting happen inside the generative model — Runway, Veo, Kling, Higgsfield — and that is exactly where they should happen. Kompozy is not a text-to-video model and does not pretend to dial in your dolly-in; the directed shot is the model's to render. What Kompozy is is the generation-and-publishing engine that wraps around those clips and does the two things a raw render cannot: it generates a great deal of net-new video the camera-control tools don't touch, and it turns any directed shot into finished, on-brand, scheduled content across every platform instead of a file in your downloads folder. The camera is the model's job; the system around it is the part that decides whether a beautiful clip ever becomes a post.
The one place motion-prompting discipline carries directly into Kompozy is the generative hook. Kompozy's Persona VFX HeyGen format prepends a short generative VFX hook — produced through fal.ai and AI-ranked against the script — to an avatar video, and that hook is a pure motion-prompting surface: the same rules from this guide (one strong move, specific and positive, chosen for impact) are what make a three-second opener land rather than drift. More broadly, Kompozy generates video the camera tools can't: Persona Shorts and longer Persona HeyGen scenes driven by a consistent avatar identity, Persona Frames that composite that avatar into a brand-exact template, Clipped Shorts that cut a long recording into verticals, Listicle and Naturalistic videos that lay animated cards over a portrait stock clip, and Marketing Shorts that composite a hook with demo footage and music. The directed clip you prompted into a model is one input among those, not the whole output.
The compounding value is on the other side of the render. A single camera-directed clip is a dead end if shipping it means manually cropping it for nine destinations, writing nine captions, and posting nine times. Kompozy fans a finished video to the eight social platforms plus blog and email, on a schedule, through a per-post review gate that keeps a human on the approve step — or, once a source has earned the trust, through Autopilot, which swaps that human gate for four automated checks (Persona Brief, platform-cadence, fact-anchor, and brand-safety) so the loop can run unattended. Either way, each persona's voice is held consistent by the Persona Brief so the on-brand framing survives the trip to every platform. So the division of labor is clean and worth internalizing: spend your craft on the camera inside the model, where this guide's vocabulary pays off, and let Kompozy absorb everything after the render — the net-new formats, the brand consistency, the multi-platform distribution — so one well-directed shot turns into a week of content instead of a single download.
Camera controls are the directions that tell a generative video model how to frame and move the virtual camera, separate from what is in the scene. They cover shot size (wide, medium, close-up), camera angle (eye-level, low, high), camera movement (pan, tilt, dolly, zoom, orbit, tracking), and lens characteristics (focal length, depth of field). In most models you supply these as words in the prompt; a few tools expose them as presets or sliders you pick directly.
Motion prompting is the practice of writing camera and subject movement into a text-to-video prompt using real cinematography vocabulary, because most models infer the camera from the prompt rather than from a dial. The reliable form is specific and singular: name one clear move — 'slow dolly-in,' 'pan right,' 'locked camera' — rather than a vague 'camera moves,' and phrase it positively, since these models generally do not honor negative instructions like 'no camera movement.'
State the move explicitly and keep it to one primary movement per shot. Use filmic terms the model recognizes — dolly in, pan right, tilt up, tracking shot, static — and phrase a held shot positively ('the camera remains still') rather than as a negation. If you want movement, you have to say so: when a prompt gives no camera direction, models like Veo tend to default to a static frame or a simple pan. For image-to-video, match the move to the image so a locked subject gets a locked camera.
Prompt-based control is the default across Runway, Veo, Kling, and Sora: you describe the move in words and the model interprets it, which is flexible but less repeatable. Preset camera moves invert that — Higgsfield's main surface is a gallery of named cinematographer moves (Dolly In, Crane Up, Bullet Time, Crash Zoom In) you select before writing the prompt, so the preset controls the camera path while the prompt controls content and timing. Presets trade some freedom for a consistent, repeatable move.
Usually one of three causes. The direction was vague — 'camera moves' gives the model nothing, where 'slow push-in' does. Too many moves were requested at once, so the model averaged them into drift; stick to one primary move per clip. Or the instruction was phrased as a negation ('no camera shake'), which these models tend to ignore — rephrase it as a positive statement ('the camera is locked and still'). Matching the move to the subject and adding a settling endpoint also stabilizes the result.
AI video camera controls are the cinematography directions — shot size, camera angle, movement, and lens — that tell a generative video model how to frame and move the camera, separate from what is in the scene. Most models (Runway, Veo, Kling, Sora) read these from the text prompt, so the working skill is motion prompting: naming one clear move at a time in filmic terms and phrasing it positively. Higgsfield inverts the approach with named preset moves you pick before describing content. Explicit, single-move, positively-worded direction produces the most reliable shots.
Get started → · ← All guides · Compare Kompozy vs other tools