
If a video model's training data doesn't separate camera movement from content, it can render the scene and ignore the shot. This article describes the dataset we built around camera movement, with its numbers, and says plainly what it has not yet been measured on.
Method. Dataset figures are from our own training run, as first described by the author. Published research on camera-motion taxonomies is cited where it bears on the design. No benchmark score is reported.
What this means for you
- Write video prompts the way a camera report reads, for example "slow dolly in, 50mm, shallow depth of field". On this model, captioning like a DP is what got shots that followed the move.
- When you test any video model, keep the content fixed and change only the move and lens. That shows you which camera terms it actually follows.
- Check zoom-in versus dolly-in on every model you try. It is the mix-up that people and automatic captioners make most often in the research cited here.
- If a model keeps giving you near-still "slideshow" shots, the likely cause is training data heavy in locked-off footage. Ask for movement explicitly and compare models on it.
- Expect identity to drift across shots, even on a model trained for camera control. Plan to check faces shot to shot.
Why camera, and not content
Most video models are trained to reproduce what is in a frame. A director mostly asks for how the frame moves: a slow push, a lateral track, a crane down. If the training data doesn't separate those two things, a model can satisfy "a cabin at dusk" perfectly and still ignore "slow dolly in, 50mm".
We wanted the reverse: a model that gets the camera right first, even when the content is ordinary. So the dataset was built around camera movement, and the content was whatever the shot happened to contain.
The dataset in numbers
| Item | Figure |
|---|---|
| Still frames | 3,859 |
| Movie clips | 1,625 |
| Camera-movement labels | every clip, labelled by hand |
| Label vocabulary | 28 camera primitives |
| Static shots | capped at 25% of the dataset |
| Clips that are pure camera motion | about 70% |
- Pure camera motion (about)70%



