Prompt-to-Motion Workflows

Motion prompting is image prompting plus time. You need everything an image prompt requires — subject, environment, lighting, style, technical specs — and then you need to describe what changes across the duration of the clip. This is where most people fail. They describe a scene and expect the model to know what should move. The model does not know. You have to tell it.

The key distinction is between camera motion and subject motion. Camera motion — pans, tilts, dollies, zooms — changes how the viewer sees a static scene. Subject motion — a person walking, water flowing, leaves blowing — changes what is happening within the scene. Most AI video models handle camera motion more reliably than subject motion because camera transforms are mathematically simpler. A slow dolly push toward a subject is almost always coherent. A person picking up an object is a coin flip.

1. Describe the Static Scene First Write your prompt as if you were generating a still image. Get the composition, lighting, and style exactly right. Then add motion as a layer on top. If the static scene does not work, no amount of motion description will save it.
2. Add Camera Motion Use cinematographic language: "slow dolly push in," "gentle pan left to right," "static wide shot," "handheld slight movement." Be specific about speed and direction. "Camera moves" is too vague. "Slow 3-second dolly forward, starting from medium shot to close-up" is usable.
3. Add Subject Motion Describe motion that is simple and physically plausible. "Hair blowing in wind," "steam rising from coffee," "clouds drifting across sky" work reliably. "Person reaching across table to pick up a glass" may produce artifacts. Simple, continuous motions outperform complex, multi-step actions.
4. Set Duration Expectations Current models produce their most coherent output in the 2-4 second range. Beyond 6 seconds, temporal consistency degrades — subjects shift appearance, lighting changes, and the overall feel drifts. Plan for short clips that you can edit together, not long single takes.