Learn the camera words that AI video models actually respond to, from a slow dolly push to an orbit around a subject and a high drone pass. Copy the prompt patterns, test them on a short clip first, and see which moves hold up and which ones models still get wrong.
Most AI video clips look generated for one reason: the camera does nothing interesting. The subject moves, the background shimmers, and the frame stays as flat as a webcam. Camera words fix that. A dolly, an orbit or a drone pass tells the model how the viewer travels through the scene, and a few precise words in the prompt are usually enough to turn a drifting clip into a deliberate shot.
A video model has two jobs in every generation: decide what moves in the scene, and decide how the camera sees it. If your prompt only covers the first job, the model fills in the second on its own, and its default is usually a gentle, aimless float. The result is technically a video and emotionally a slideshow.
Film language already solved this. Directors do not say make it feel closer, they say push in, and every crew member knows what that means. Video models trained on captioned footage picked up the same vocabulary, so the old terms for dolly, truck, pan, tilt, pedestal and orbit work as steering words. Write them like a shot list and you get something closer to a shot.
The practical payoff is on short clips. Picasso IA lists 121 text-to-video models and 28 video-editing models in the full catalog, and a clip from any of them is only a few seconds long, so there is no room to wander. One clear move per clip is the difference between footage you can cut into an edit and footage you delete.
Before you write anything, get the vocabulary straight, because half of the failures come from using a word that means something else on a set. A dolly is the whole camera travelling forward or back. A zoom only changes the lens and leaves the camera where it is. The table below maps each term to what you should see and the wording that tends to read well.
| Camera term | What happens in frame | Prompt wording to try | Watch for |
|---|---|---|---|
| Dolly in | Camera glides toward the subject, background parallax shifts | slow dolly in toward her face, steady, smooth | Subject and background growing together like a zoom |
| Truck sideways | Camera slides left or right, foreground passes faster than background | camera trucks left along the shopfronts | Model panning in place instead of travelling |
| Orbit | Camera circles the subject while keeping it centered | slow orbit around the statue, one half circle | Subject rotating while the camera stays still |
| Pan and tilt | Camera turns left, right, up or down from a fixed point | pan right across the harbor, then hold | Fast pans smearing detail |
| Pedestal or crane | Camera rises or drops vertically | crane up from the street to reveal the rooftops | A drifting zoom instead of a real lift |
| Drone aerial | High forward or reverse flight, often a reveal | drone shot flying forward over the pine forest | Impossible flight paths and a shaky horizon |
Two terms in that table cause most of the confusion. Orbit and drone sound unambiguous but are the ones models mix up most, so they get the most space in the next sections.
This is the routine that keeps the camera from being an afterthought. It works the same way on any video model in the toolkit.
Open the toolkit, choose a video model and keep a text file of the winners.
Treat these as patterns, not magic phrases. Swap the subject and keep the structure: a scene sentence, then a move sentence, then the speed and the end point.
For a dolly, try: a weathered lighthouse keeper stands at a window at dusk, the camera dollies in slowly until his face fills the frame. For an orbit: a ceramic teapot on a wooden table, the camera orbits slowly around it in a half circle, soft window light. For a drone shot: drone shot flying forward low over a turquoise coastline toward a cliff village, golden hour.
Different models take these differently. The video-01-director model is built around camera control: you put tags in square brackets, such as [Pan left] or [Zoom in], and you can combine up to three movements in one prompt, for example [Truck left, Pan right, Zoom in]. Its listed moves include truck, pan, push in, pull out, pedestal, tilt, zoom, shake, tracking shot and static shot. It has no orbit tag, so an orbit there is a plain-language experiment.
Most other models, among them kling-v3-video, veo-3.1, seedance-2.0, wan-2.6-i2v, hailuo-2.3 and ray-2-720p, have no tag syntax and read camera words as part of an ordinary sentence. kling-v3-video also lets you script up to six shots in one generation, which suits a sequence where shot one is a dolly and shot two is a drone reveal. For starting from a still, the guide to turning photos into videos covers the image-to-video route, where you attach the photo and spend the whole prompt on the move.
The first mistake is stacking. Asking for a dolly in, an orbit and a tilt up in one short clip gives the model three instructions and very little time, so it usually picks one and blends the others into mush.
The second is using lens words as camera words. Zoom in and dolly in are different shots, and on many models both produce a similar push. If you care about parallax, the sense that the background slides behind the subject, say the camera moves and name the direction instead of only saying zoom.
The third is skipping the end point. A move without a destination just keeps going until the clip stops. Saying until the face fills the frame or ending on the rooftops gives the model something to aim at.
Tip: if a move keeps failing, shorten it. Replace a full orbit with a quarter turn, or a long drone flight with a short forward push. Small moves are far more likely to come out clean, and you can always join two clean clips in an edit.
Camera prompting is useful, but it is not a camera. These are the places it stops delivering, so you can plan around them.
A dolly moves the whole camera toward or away from the subject, so the background shifts at a different rate than the foreground. A zoom changes the lens and leaves the camera in place, so everything scales together. In an AI prompt, say dolly in or the camera moves forward when you want depth and parallax, and zoom in when you want the flat optical push. Many models blur the two, so run a short test to see which one your model separates.
Describe the subject first, then add a separate camera sentence such as the camera orbits slowly around it in a half circle. Name the direction, the speed and where the move ends. Ask for a partial arc rather than a full circle, because a complete orbit exposes the back of the subject that the model has to invent. If the subject spins instead of the camera, reword it to the camera circles around while the subject stays still.
There is no single answer, and the catalog is the reason to test. The video-01-director model is the one built around camera control, with bracketed tags for pan, tilt, truck, push, pedestal and zoom. Others such as kling-v3-video, veo-3.1, seedance-2.0, wan-2.6-i2v, hailuo-2.3 and ray-2-720p read camera wording from plain sentences. Run the same move prompt through two or three of them on a short clip and keep whichever follows it most closely.
Sometimes. The video-01-director model accepts up to three movements in one bracket, for example a truck left with a pan right. On models without tags, two moves in one sentence can work if they are compatible, such as a dolly in with a slight tilt up, but opposing or competing moves usually collapse into one. When it matters, split the idea into two shots and join them in an edit, or use a multi-shot mode where the model offers one.
Start with the word drone and a direction: drone shot flying forward over a pine forest at sunrise. Add altitude with low or high, and the reveal if you want one, such as rising to show the valley behind the ridge. Keep the path open, with no tight gaps and no crowds, because the model is imitating the look of aerial footage, not simulating a flight. A steady horizon is the first thing to check in the result.
Yes, and it is often more reliable, because the first frame is fixed and the prompt only has to describe motion. Attach the photo, then spend the whole prompt on the move and what changes in the scene: the camera dollies in slowly while the curtains move in the breeze. Keep the move modest, since a big camera travel forces the model to invent areas of the scene the photo never showed.
Usually because the wording was ambiguous or the move was too ambitious for the clip length. Models treat orbit, dolly and zoom as near neighbors. Try rephrasing in plain language such as the camera physically moves around the subject, remove any competing instruction, and shorten the move. If the same prompt keeps failing on one model, that model may simply not do that move, and another one in the catalog might.
As short as the model allows. A test exists to answer one question, did the camera do what I wrote, and a few seconds is enough to see it. Longer clips cost more credits and give the model more room to wander, so wait until the move is working before you extend. Check the model page for its available durations, and compare costs on the pricing page before you plan a long take.
The prompts are plain words, so they are yours to use. What you can publish depends on the footage: check the commercial terms that apply to the model and to your plan, and keep the subject something you have the right to show. For product footage, make sure the camera move does not distort the item or hide its real shape. For real people, only use likenesses you own or that are fictional.
No. Camera words are just part of the prompt, so they add no separate charge. Generation costs credits, and the amount depends on the model and the settings you choose, such as duration and resolution. New accounts get free credits to test a few clips. Current plan details are on the pricing page, which is the only place worth trusting for numbers.
Ready to try your first move? Open the Picasso IA toolkit, pick a video model and run a short dolly test on your own scene.
Choose your language
