image to video motion planning: Fit motion complexity to the reference and clip

Good image to video motion planning starts with what the reference actually shows. A still image can establish appearance and composition, but it does not supply every hidden surface or every intermediate pose. Limit the requested motion to a reviewable objective and make uncertainty visible.

Inspect the reference before writing action

Identify the subject, visible components, crop boundaries, likely motion direction, and any parts already hidden. Note whether the source cuts off hands, wheels, or object edges that the planned movement would reveal.

A fictional product photograph showing only the front of a bottle may be suitable for a small camera move or restrained environmental motion. A full rotation would require the output to invent unseen sides. That can be useful for a concept, but it is not evidence of the actual product’s rear geometry.

Choose the primary reference deliberately. The current adapters normally forward only the first imageInputs entry as image_url. Additional uploaded views do not automatically reduce uncertainty through multi-image fusion.

Allocate a motion budget

List the changes the clip asks the viewer to follow: subject action, camera movement, background change, occlusion, and newly revealed surfaces. This is a qualitative planning tool, not a measured model capacity score.

For an initial concept, reduce the number of simultaneous changes. A proposed prompt might keep the product stationary while a soft curtain moves gently in the background and the camera makes a restrained forward move. Another concept might hold the camera still and ask for one simple subject action.

Do not assign an invented maximum number of motions to a model. Use your task’s acceptance criteria and observed results to decide whether the request is too complex.

Match the actual duration

Set a valid duration in the request body rather than relying on narrative timing. The current schemas allow Seedance integer seconds from three through fifteen, Kling five or ten, and Veo eight.

A longer allowed duration is not a guarantee of better motion or fidelity. It simply changes the requested configuration and, depending on the model, the credit calculation. Review the full request and budget before increasing it.

Inspect the resulting changes

Compare the output with the approved reference at the start and through key transitions. Check identity, geometry, label placement, and any newly visible region. Watch for changes hidden by motion blur or occlusion.

If the final crop removes the reference’s strongest identity cues, reassess the delivery rather than approving only the uncropped source.

Accept the clip when its visible motion meets the brief and invented details remain within the approved concept boundary. Continue with reference fidelity testing, motion continuity review, and multiple reference handling before expanding the motion plan.

Continue with quality and testing

• generation model regression checklist: Maintain a release scorecard with real evidence
• generation prompt revision diff: Review the change as a field-level hypothesis
• AI image color drift test: Control viewing conditions before judging color
• Browse all quality and testing guides