AI video with camera control — 17 moves, lenses and lighting

Two AI video tools can produce identically pretty frames and still be wildly different products, and the difference is almost always camera control. One lets you describe a subject. The other lets you describe a shot. Once you have worked with the second kind, the first feels like directing a film by mailing the actors a description of the room and hoping.

Here is what camera control actually means, which controls exist on eroq, what engines can and cannot follow, and how to write prompts that land.

Three levels of "camera control"

Level one — you type it and hope. The model reads "dolly in" as one more adjective competing with everything else in the sentence. Sometimes it works. You have no way of knowing which word did it.

Level two — named controls folded into the prompt. You pick a move from a list, and the platform composes it into the engine prompt in the phrasing that engine responds to. You still don't control the camera directly, but you control the instruction precisely, and it is the same instruction every time — which is what makes iteration possible.

Level three — an actual 3D camera path. Keyframed position and focal length, resolved before the frames are generated. Almost nothing in consumer AI video does this today. When a vendor claims it, ask what happens to the subject when the path moves, and verify on their site as of September 2026.

eroq is level two, and says so. The camera controls are named, the composition happens server-side, and written directions in your own prose are detected too — type "wide shot", "anamorphic" or "f/1.4" in the paragraph and they are read as direction, not decoration.

The 17 moves, grouped by what they're for

• Hold the frame. Static — the camera does not move. Underrated. If the subject is doing the work, a locked frame is the strongest choice, and it is the least likely to warp.
• Reveal slowly. Slow pan, Push-in, Pull-back, Crane up, Crane down. These change what the audience knows. A push-in commits to a face; a pull-back admits context you were hiding.
• Follow. Tracking, Handheld, POV, Snorricam. Tracking rides alongside, handheld adds human unsteadiness, POV puts the lens in the subject's eyes, Snorricam rigs the camera to the body so the world lurches and the face stays still.
• Circle. Orbit. One subject, one continuous arc. The most reliable way to look expensive.
• Fly. FPV drone, Aerial pull-back. Speed and altitude, best given something to pass close to.
• Frame. Close-up, Wide. Not moves at all, but the two you will use most.
• Break the rules on purpose. Dolly zoom (the frame stays, the background rushes) and Whip pan (a violent smear, usually a cut in disguise).

Seventeen names, and the useful discipline is one move per clip. Two moves in a 5-second shot is how you get mush.

The rest of the rack

A move is direction. The rest of the look is the rack, and it is folded into the prompt the same way:

• Camera gear — 35mm film, IMAX, VHS camcorder, Drone, Action cam, CCTV, Smartphone. This one does the heaviest lifting of any single control, because it sets grain, aspect feel and the implied competence of the operator.
• Lens — Clean sharp, Anamorphic, Vintage anamorphic, Soft portrait, Macro, Fisheye.
• Aperture — f/1.4, f/4, f/11. Depth of field, stated as a number the model recognizes.
• Lighting — Silhouette, Practicals, Window light, Overhead fall, Contre-jour, Soft cross, Neon glow.
• Palette — twelve of them, from Neon noir and Candy pop to Bleach bypass and Blood & gold.
• Film type, era and tempo — Film noir through Romance, 1920s through Futuristic, Calm through Chaotic.

Pick four or five, not all of them. A prompt carrying every control at once is a prompt with no priorities, and the engine will average them.

What engines can and cannot follow

Honest version: every one of these is a suggestion, weighted differently by each engine, and none of them is a guarantee. What varies is which suggestions stick.

• Framing and gear stick almost everywhere. Close-up, wide, 35mm, VHS — these change the output reliably on every engine.
• Large translations are harder than rotations. Orbit and pan land more often than a long tracking move, because the model has to invent more of the world as the camera travels.
• Dolly zoom is the least reliable move on the list, on every engine. It requires two things to change in opposite directions at once. Expect to spend takes.
• Faces under motion are where Kling 2.5 Turbo earns its credits; physics and hard light are where Hailuo 02 does.
• Reproducibility belongs to the Seedance family, which honors a seed. Seedance 2.5 adds first/last-frame interpolation, which is the closest thing on the roster to guaranteeing where a move ends up.
• Shorter clips follow direction better. A move has to survive every frame; 5 seconds gives it fewer chances to drift than 15.

Three worked examples

Orbit, on a single subject.

A welder lowers her mask and strikes an arc against a steel beam in a half-built hangar, sparks falling in sheets around her boots. Slow orbit around her at waist height, IMAX, clean sharp lens at f/4, contre-jour light through the open end of the hangar, industrial fog palette, tense tempo.

Push-in, on a face.

A chess player sits alone in an emptied club long after the last game, one hand resting on a toppled king while the radiator ticks behind him. Slow push-in to a close-up, 35mm film, soft portrait lens at f/1.4, window light falling from the left, nostalgic blue palette, calm tempo.

FPV drone, for speed.

A drone chases a motorcycle down a switchback road cut into a dry hillside, dust boiling off the rear tire as the rider leans into each turn. FPV drone shot, action cam, clean sharp lens at f/11, hard overhead sun, warm film palette, chaotic tempo.

Each names subject, motion, camera, light and mood, in one paragraph of 30 to 80 words, with exactly one move. Load the first one straight into the tool: /studio/video?prompt=A%20welder%20lowers%20her%20mask%20and%20strikes%20an%20arc%20against%20a%20steel%20beam.

When a move doesn't land

Work through these in order, changing one thing at a time:

1. Shorten the clip. Five seconds instead of ten. Most drift is duration.
2. Remove competing motion. If the subject is also running, the camera move has to share the frame budget. Let one of them move.
3. Pin a seed on a Seedance engine and change only the move. Now you are comparing direction, not luck.
4. Add a negative prompt on an engine that supports one — "shaky, warped hands, jump cut" removes more than adding adjectives ever will.
5. Say it twice, once as a control and once in your prose. Redundancy is cheap and it biases the weighting.
6. Change engine. Some moves simply land better elsewhere, and a 60-credit test on a cheap engine answers the question faster than a fourth take on an expensive one.

FAQ

How many camera moves does eroq support?

Seventeen named moves, from Static and Slow pan through Orbit, Dolly zoom, FPV drone and Snorricam, plus written directions detected in your own prompt text. They combine with camera gear, lens, aperture, lighting, palette, film type, era and tempo.

Do camera controls work on every video engine?

They are folded into the prompt, so every engine receives the direction — but each interprets it differently, and none guarantees it. Framing and gear cues land most reliably, long tracking moves and dolly zooms least.

Can I get the exact same shot twice?

On the Seedance engines, yes within reason — they honor a reproducibility seed, and Seedance 2.5 supports first/last-frame interpolation. Pin the seed, change one control, compare.

Pick a move and shoot it — /studio/video.