Most failed AI clips fail before the video tool ever runs: they fail at the still. Motion amplifies whatever a frame already is — a strong composition becomes cinema, a muddled one becomes soup. This workflow is the studio’s standard route from a gallery of stills to clips that hold together.
Step one: choose like an editor
Pick stills with one subject, a readable pose and a background simpler than the subject. Check the hands and the hairline at full size; motion finds both. Frontal and three-quarter poses animate most reliably. If a still needs an apology, it needs a re-render, not a video job.
Step two: direct one movement
Write the motion brief with a single verb: turns, steps, rises, glances. Add a manner (“slowly”), a camera instruction (“camera still”) and stop. Stacking verbs — “she turns, smiles, walks and waves” — gives the model four jobs and it will do half of each. Small motion, completely done, reads as confidence on screen.
Step three: test at the shortest length
Price follows duration and resolution, so the cheap test is also the honest one: prove the motion short. Watch the face through the fastest movement and the background at the edges. If it holds, extend; if it tears, the fix is almost always a smaller motion or a simpler still, not a longer render.
Step four: finish in order
Face work before upscaling, upscaling dead last. Recast with the video face swap while the clip is cheap to redo, lock the cut, and only then spend on resolution. And keep the original still and clip untouched in the library — every stage a sibling, never an overwrite — so any step can be revisited without starting the shoot again.