Ask five creators which AI video model has the best camera control and you will get five different answers, because "camera control" is not one capability. It is at least five: understanding a movement described in text, executing named presets, following a reference video's motion, holding a composition you defined, and keeping a whole multi-shot sequence coherent while all of that happens. A model can nail "pan left" and still melt your product's shape by shot three.
A current film-focused leaderboard with a dedicated camera-and-lighting criterion gives the debate some numbers. Here is what they say, where they stop short, and how to write prompts that make any model follow direction better.
The camera-and-lighting board
| Model | Camera & lighting score | Position | Listed cost per 5s |
|---|---|---|---|
| Seedance 2.5 | 1,222 | 1st | $1.16 (720p) |
| Seedance 2.0 | 1,160 | 2nd | $0.76 |
| Wan 3.0 | 1,135 | 3rd | $0.50 |
| Google Omni Flash | 1,128 | 4th | $0.50 |
| Kling 3.0 Omni | 1,027 | Joint 9th | $0.42 (no audio) |
The headline finding deserves emphasis because it cuts against how these tools are marketed: a detailed camera vocabulary does not guarantee strong execution. The model with the most thoroughly documented movement terminology — Kling, which publishes named support for push-ins, pull-backs, pans, tilts, forward tracking, slow orbits and static shots — sits near the bottom of the judged results. The leaderboard scores what the clips look like, blind, not what the feature list promises.
The method: identical prompts and settings per model, one output each entering judging under a fixed rule, model names hidden, positions randomized, forced pairwise choices converted through Bradley-Terry scoring, with confidence intervals published. The 62-point gap from first to second is worth checking against those intervals; the 195-point gap to Kling is the separation that actually matters.
What each score does and does not tell you
The camera-and-lighting criterion bundles two things — how the virtual camera behaves and how the scene is lit — and it does not crown separate winners for pans versus tracking shots versus dolly zooms. Fair comparison on this kind of board also separates several sub-metrics that usually get blurred together: instruction adherence (did it attempt the move you asked for, without inventing another), path precision (direction, speed, distance), subject stability (do faces and products hold form while the camera moves), temporal consistency (flicker, frame-to-frame drift), and physical realism — the one that distinguishes a real dolly, where perspective and parallax change, from a cheap zoom that just enlarges the frame.
Orbits are the stress test with public evidence: the leaderboard's signature camera prompt runs a fast 360-degree orbit around a seated subject across every model, and it is exactly where identity and geometry errors surface first, because the model must invent the hidden sides of a face. Tracking shots compound difficulty by adding subject motion on top of camera motion. For pans, tilts, zooms and crane moves, no public per-move ranking exists yet — those still require your own same-prompt tests.
Prompt structures that survive contact with reality
Vendor guidance from the most documentation-heavy model points to the same conclusion across the field: prompts fail when they stack multiple moves, contradict themselves, pit fast subject action against the camera path, or lean on vague words like "cinematic." One dominant move, with direction, speed and stability constraints stated, works better. Four reusable structures:
- Subject + move + purpose. "A ceramic watch rests on a stone pedestal. The camera slowly pushes in to emphasize the engraved dial." Best when the movement serves a specific visual job.
- Start frame + end frame. "Start on a wide view of the chef and kitchen. End on a close view of the knife cutting pastry. The camera physically moves forward while focal length stays constant." Naming the framing change — and the constant focal length — is what separates a dolly from a zoom in the model's interpretation.
- Move + speed + stability. "Pan right slowly at constant speed while the cyclist stays centered. Keep the horizon level; no zoom, no rotation." Use when unwanted drift is the main enemy; exclusions are constraints too.
- Shot list for sequences. "Shot 1: wide static view of the lobby. Shot 2: slow track behind the guest toward reception. Shot 3: tilt down to the room key. Preserve clothing, lighting and screen direction across all shots." One dominant movement per shot, identical subject and scene details reused throughout.
Matching model to job
Narrative and film work: Seedance 2.5 carries the strongest judged evidence — it leads the camera criterion and the broader film and overall boards, so the same case for it covers cinematic quality generally.
Volume and social content: iteration speed and usable yield beat a few ranking points. At $0.42 and $0.50 per five seconds, Kling 3.0 Omni and Wan 3.0 buy you several attempts for the price of one Seedance 2.5 clip. That is also the honest way to price them: cost per usable shot means dividing total spend — retries and rejects included — by clips you would actually publish. Wan 3.0's third-place score at mid-pack price makes it the value pick when quality still matters; Kling's price advantage is partly eaten by its weaker judged execution.
Multi-shot planning: if your workflow genuinely needs per-shot camera direction with reference-based continuity — the feature Kling documents as storyboard controls with per-shot duration, size, perspective and movement — its vocabulary may still earn its place even at a lower execution score. Feature documentation is a workflow signal; leaderboard scores are a quality signal. Use both.
The practical verdict
Seedance 2.5 is the current leader for judged camera and lighting execution, with Seedance 2.0 as the cheaper sibling and Wan 3.0 as the volume value. Kling 3.0 Omni is the planner's tool with the cheapest clips and the richest named-movement vocabulary, but the board shows its execution trails. And no leaderboard replaces the final step that every serious pipeline uses: lock one prompt, one product, one camera path, run it across your two or three candidates with a fixed pass bar, and count what survives. The camera move you can afford to test is the one that actually tells you which model controls the camera — yours.
