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 references (and use 2.5's multimodal inputs)
For image-to-video, the reference already supplies subject appearance, composition, color, and often lighting — so don't waste the prompt redescribing them. Direct what a still cannot show: movement, camera, timing, environmental response, sound, and the final frame.
Seedance 2.5 expands what a "reference" can be: one generation can read up to 30 images, 10 video clips, and 10 audio clips at once, so multi-subject scenes (a band, a crowd, a product line) can pull each subject's appearance and sound from its own source. Source ↗
- Use a clean subject reference when identity or product shape matters.
- Avoid references that disagree on clothing, lighting direction, or proportions.
- For a camera move, leave visual space in the direction the frame needs to travel.
- White-model reference (new in 2.5): block in spatial structure, subject pose, motion path, and camera placement with an untextured 3D model first, then let the model render style on top. Use it when a complex shot's composition and blocking matter more than its surface look.
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.
Original concept loops · hover any frame to pause
Use timed beats when the action changes
With 2.5's 30-second takes, timed beats are how you keep a longer shot readable — but don't fill every second with a new event. A four-beat structure is often enough: establish, act, reveal, resolve. Give the final beat 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.
| Failure | First edit to try |
|---|---|
| Subject appearance drifts | Strengthen the reference and remove conflicting appearance adjectives. |
| Camera ignores direction | Use one move, one height, and one endpoint. |
| Action feels rushed | Remove secondary actions or split the sequence into beats. |
| Objects appear or vanish | Name the important object and state where it remains at the end. |
| Sound feels generic | Name two or three diegetic sounds and omit vague music direction. |
6. Run the five-point QA pass
- Identity: face, clothing, product geometry, and text remain stable.
- Physics: hands, hinges, liquids, shadows, and contact points behave plausibly.
- Continuity: important objects persist and spatial relationships remain readable.
- Camera: the path is smooth, motivated, and ends on the intended frame.
- 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.
7. Extend a clip across multiple rounds
Seedance 2.5 can run a single take up to 30 seconds and extend it across multiple rounds while subject, scene, lighting, and sound stay consistent — so a long story becomes one continuation chain, not many separate clips to stitch.
Write an extension as a continuation beat, not a new scene: reference the previous clip, state exactly what must carry over, then direct only the next action.
Extend @clip1. Continue the same subject, location, wardrobe, and visual style into the next 30 seconds; keep lighting and sound consistent. The boy with the football sprints down the train aisle; when the doors open he bolts onto the platform, the man chasing him through the crowd until he catches the boy's arm. Track beside them at a run, then settle the frame as the tension releases.Keep these stable, in priority order: subject identity → location and wardrobe → lighting and palette → sound design. If continuity breaks, strengthen the reference or the "keep consistent" clause before rewriting the action.
8. Edit a specific moment by timestamp
2.5's editing tools let you revise a chosen second of a clip — swap a green-screen background while keeping the subject and its physics, re-direct the camera path, or change a reference — without regenerating from scratch.
Direct the change and its time window, and state what must stay fixed.
Edit @clip1 between 0:04 and 0:10. Keep the subject's identity, motion, and clothing unchanged. Replace the background: move from an empty training room to a stadium pitch with original defenders and a goalkeeper; the subject dribbles past two defenders and shoots. Render the new scene's physics on the subject — shirt, hair, and stride respond to the new ground and light.Green-screen edits work best with an evenly lit subject and a clean key; viewpoint edits work best when you name one new camera behavior and its endpoint instead of stacking moves.
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.