The Seedance Prompt Guide: Structure, Examples & Best Practices

A great Seedance clip almost always starts with a well-built prompt. The good news: prompting is a skill, not luck. This guide breaks down the structure that consistently produces usable video, with a reusable formula and worked examples. (For the quick version, see how to use Seedance; this is the deeper one.)

The anatomy of a Seedance prompt

Strong prompts stack five layers. You don't need all five every time, but naming each one removes guesswork:

1. Subject & action — who or what is in frame, and what they're doing. "A lone astronaut planting a flag," not just "an astronaut."
2. Camera & motion — give the camera a job: "slow dolly in," "aerial drone descending," "360° turntable," "handheld follow." Static descriptions produce static-feeling clips.
3. Lighting & mood — "golden hour," "hard rim light," "soft studio key," "moody low-key." Light is what makes video read as cinematic.
4. Style & finish — "photoreal, anamorphic lens flares, film grain," or "2D hand-painted anime," or "clean product render." This locks the look.
5. Audio — Seedance 2.0 generates native audio, so describe it: "ambient city rain," "soft piano," "a calm voice saying…" for dialogue and lip sync. Most people skip this and leave half the model's power unused.

A reusable formula

[Subject + action] + [camera & motion] + [lighting & mood] + [style & finish] + [audio]

Keep it one focused idea per prompt. Longer isn't better — a bloated prompt with conflicting instructions gives the model too much to reconcile.

From basic to great

Watch the same idea improve as each layer is added:

• Basic: A woman walking down a street.
• Better (＋camera ＋lighting): A woman walking down a rain-slicked city street at night, neon signs reflecting on the pavement, slow tracking shot beside her.
• Best (full five layers): Cinematic tracking shot following a woman in a red coat down a rain-slicked Tokyo street at night; glowing neon reflects on wet pavement; shallow depth of field, anamorphic lens flares, moody color grade; ambient rain and distant traffic.

The third version isn't longer for the sake of it — every clause does a specific job. Browse the prompt library to see dozens of real examples built this way.

Text-to-video vs image-to-video prompts

The structure shifts depending on your starting point:

• Text-to-video (t2v): you describe the whole scene, so lead with subject and setting, then camera and style. You own the first frame, so be specific about composition.
• Image-to-video (i2v): the image already sets the look — your prompt is mostly about motion. Keep it subtle ("gentle head turn," "hair moving in the breeze," "camera slowly pushes in") and add "preserve identity" to avoid face drift. Over-describing a still you've already provided fights the source image.

Best practices

• Always name the motion. It's the single biggest upgrade for most prompts.
• Change one thing at a time. Iterate camera, then lighting, then style — so you learn what each edit does.
• Match aspect ratio to the platform — 16:9 landscape, 9:16 vertical, 1:1 for product loops.
• Stay focused. A tight, well-ordered prompt beats a long, contradictory one.

Common mistakes

• Vague subject ("a nice video") — the model has nothing to anchor to.
• No camera direction — you get a flat, lifeless shot.
• Conflicting styles ("photoreal 2D anime cinematic oil painting") — pick a lane.
• Ignoring audio — on Seedance 2.0 that's a wasted dimension.

Try it

The fastest way to internalize this is to generate. Copy a proven prompt from the library, run it on ponpon, then change one layer and re-run to see the effect. For what the model can do with these prompts, see Seedance 2.0 features.