Models trained on film respond to film language. "Dolly in slowly on the subject" produces a repeatable result; "make it cinematic and emotional" produces a lottery. The gain is not aesthetic, it is variance: a named move gives you a shot you can ask for twice.
The moves that work are the ones a crew could physically do, plus a small number that only exist because the camera is imaginary: the unbroken orbit-to-street zoom, the fly-through an aperture too small for a rig. Those are worth reaching for precisely because a shoot could not deliver them.
We publish our working reference of thirty-eight camera movements with a usable prompt for each, free and without an email gate.
Which camera moves work best in AI video generation?
Single, clearly named moves with one intention: slow dolly in, truck left with parallax, crane down, rack focus. Compound moves (orbit while craning while pulling focus) fail more often, because the model has to satisfy three constraints at once.
Why does my AI video ignore the camera direction?
Usually because the prompt describes mood alongside movement, and the mood words dominate. Put the move first, in camera vocabulary, in its own sentence, then describe the subject.