Before you startAccess, model names, clip length, resolution, and reference limits can differ across regions and products. Confirm the available options in your own interface. This guide focuses on the directing workflow that remains useful across surfaces.

1. Define one shot objective

Write one sentence that describes what the viewer must see happen. Good objectives have a beginning and an end: “Start on the unopened box, then reveal the product as the lid lifts.” Weak objectives describe taste instead of action: “Make a luxurious product video.”

Keep the first attempt to one shot. If you need a sequence, divide it into beats before prompting. Each beat should have one dominant subject action and one camera intention.

2. Prepare only useful references

For image-to-video, the image already supplies subject appearance, composition, color, and often lighting. Do not waste the prompt redescribing every visible detail. Direct what the still image cannot show: movement, camera, timing, environmental response, sound, and the final frame.

  • Use a clean subject reference when identity or product shape matters.
  • Avoid references that disagree on clothing, lighting direction, or proportions.
  • Crop out irrelevant objects that the model may try to preserve.
  • For a camera move, leave visual space in the direction the frame needs to travel.

3. Write the prompt in production order

Use this order: subject → action → camera → composition → setting → light → style → audio → endpoint. It reads like a compact call sheet and makes debugging easier because every phrase has one job.

A red folding bicycle stands beside a rain-streaked café window. A hand enters frame, unfolds the bike, and locks the hinge. The camera tracks left at waist height, moving from a medium side view to a close detail of the locking mechanism. Early-morning street, wet pavement reflecting warm interior light. Clean commercial realism. Audio: soft rain, hinge click, distant traffic. End with the bike fully open and the lock centered in frame.

Camera moves, seen in motion

Six moves cover most first-pass direction. Watch the frame, not the subject — the camera decision is what changes between loops, and each one implies a different endpoint. Pull-back is simply push-in run in reverse. These are original concept loops drawn for this guide, not model output.

REC00:00–00:06
Locked-off staticThe camera holds. The subject's action carries the whole shot.
REC00:00–00:07
Push-inScale 100% → 145%. Earn the move by ending on a detail.
REC00:00–00:07
OrbitForeground and background slide at different rates around the subject.
REC00:00–00:07
PanOne axis, one speed, one reason to move.
REC00:00–00:07
TrackingThe camera matches the subject's pace; the world drifts behind.
REC00:00–00:03
HandheldControlled jitter. Documentary texture, not chaos.

Original concept loops · hover any frame to pause

Use timed beats when the action changes

Timed beats are useful when your surface supports longer outputs, but do not fill every second with a new event. A four-beat structure can be enough: establish, act, reveal, resolve. The final beat should give the motion somewhere to settle.

4. Choose settings for the test

Use the shortest practical clip and a moderate output size while learning the prompt. Keep the aspect ratio tied to the delivery surface: vertical for phone-first social, landscape for presentations and YouTube, square only when the composition is designed for it.

If a seed or variation control is available, keep it fixed while changing prompt language. If it is not available, compare several outputs before concluding a wording change caused the difference.

5. Generate, then change one variable

Do not rewrite the whole prompt after a near miss. Identify the first visible failure. If the subject changes identity before the camera moves, fix consistency before adding more camera detail. If the action is correct but framing drifts, simplify the move and state the endpoint.

FailureFirst edit to try
Subject appearance driftsStrengthen the reference and remove conflicting appearance adjectives.
Camera ignores directionUse one move, one height, and one endpoint.
Action feels rushedRemove secondary actions or split the sequence into beats.
Objects appear or vanishName the important object and state where it remains at the end.
Sound feels genericName two or three diegetic sounds and omit vague music direction.

6. Run the five-point QA pass

  1. Identity: face, clothing, product geometry, and text remain stable.
  2. Physics: hands, hinges, liquids, shadows, and contact points behave plausibly.
  3. Continuity: important objects persist and spatial relationships remain readable.
  4. Camera: the path is smooth, motivated, and ends on the intended frame.
  5. Audio: events line up with sound and the mix supports rather than masks the action.

Save the prompt alongside the output and record what changed. A small prompt log is more valuable than a folder of attractive clips you cannot reproduce.

Sample output

These clips show the directing techniques from this guide applied to a real render. They are currently Seedance 2.0 samples — the camera, action, and endpoint decisions transfer directly to 2.5. Each clip is badged with its source version.

Where to go next

Use the prompt template builder to assemble your own call sheet, or browse Seedance 2.5 prompt examples to see how the same structure changes across product, character, social, and cinematic shots.