How to write prompts that Seedance 2.0 Mini and Fast follow: three layers, a timeline for clips with more than one beat, image references with a job each, one line of exclusions, and a library of camera moves to copy and fill in.
Brief the model the way a director briefs a crew: what is in frame and what it does, then how the camera moves, then how the shot is lit. One or two short sentences per layer.
A barista pours steamed milk into a glass of espresso, the foam folding into a leaf. Slow push-in at counter height, 50mm, shallow focus. Warm window light from the left, soft steam, muted browns, fine grain.
Name the subject plainly, then give it one clear verb: pours, turns, lifts, walks toward. Put it first; the model weighs the opening words most.
One move per shot, with a speed, a height and a lens if it matters. "Slow push-in at eye level, 85mm" beats "cinematic camera".
Say where the light comes from and what it looks like, then the palette and texture. "Hard side light from the right" is something the model can draw; "moody" is not.
Seedance 2.0 can cut and change action inside one clip. Mark each beat with a time range so the action is spread over the length you pay for.
Tell the model what every image is for. In the API the first frame goes in image_url and up to 9 references in reference_image_urls. In the prompt they are @image1, @image2 and so on, numbered in the order you add them.
Keep the face, hair and jacket from @image1.The bottle and its label from @image2 stay exactly as shown.Match the color and light of @image3.Set the scene in the kitchen from @image4.End the prompt with a single "Avoid:" line naming the few things that tend to go wrong for this shot. A long list of negatives dilutes everything before it.
Avoid: text overlays, extra fingers, the logo changing shape.no ugly, no blurry, worst quality, low quality, bad hands, deformed, extra limbs, watermark, bad anatomyWhere you can, say what you want instead: "static camera" works better than "no camera movement", "clean white background" better than "no clutter".
The same idea, written so the model has something to render. Copy the strong version and change the details.
WeakA cool ad for headphones.
StrongMatte black headphones on a slowly turning stone plinth, a thin line of light running along the headband. Slow orbit left at product height, 85mm, deep black background. Hard rim light from behind, soft fill from the front, clean reflections. Avoid: text, the logo changing shape.
Why it worksIt names the object, one move, the lens and where the light comes from.
WeakGirl talking about skincare.
StrongA woman in her twenties holds a serum bottle up to the camera and says: "It sinks in fast and my skin feels calm." Handheld selfie framing at arm's length, slight natural sway. Soft daylight from a window, bathroom behind her out of focus.
Why it worksThe line of dialogue is in quotes, and the framing reads as a real phone video.
WeakSad scene in the rain.
StrongA man in a gray coat stops under a bus shelter in heavy rain, reads a message on his phone, then lowers it slowly. Static medium shot through rain-streaked glass, then a slow push-in to his face. Cold blue street light, a warm shop sign behind him.
Why it worksThe emotion comes from actions and light the model can show, not from the word sad.
WeakCity at night.
StrongRain-soaked crossing at blue hour, taxis and umbrellas passing, traffic lights reflected in the puddles. Locked-off wide shot from a raised angle, 24mm. Neon and headlights, long reflections, light fog.
Why it worksA place, a time of day, a fixed camera and specific light sources.
Copy a line, replace the [bracketed] parts, and drop it into the camera layer of your prompt. One move per shot works best.
Use it for drawing attention to a detail or a reaction
Slow push-in toward [subject] at [height] height, [lens], ending on [detail].
Use it for revealing the setting or closing a scene
Slow pull-out from [detail] to reveal [setting], camera at [height] height.
Use it for showing a product or character from every side
[speed] orbit [direction] around [subject], keeping it centered, [lens], [setting] behind.
Use it for following action across the frame
Camera trucks [direction] alongside [subject] as it moves, matching its speed, [shot size].
Use it for keeping a moving subject in frame
Tracking shot following [subject] from behind at a steady distance as it moves through [setting].
Use it for opening a scene or ending on a wide view
Crane up from [detail] to a high wide view of [setting], [speed].
Use it for scanning a place or linking two subjects
[speed] pan [direction] from [subject] to [detail], camera fixed in place.
Use it for revealing height or a tall product
Tilt [direction] from [detail] to [subject], slow and steady.
Use it for an energetic switch between two moments
Whip pan [direction] from [subject] to [detail] with motion blur, then settle.
Use it for moving attention between two planes
Static [shot size], focus racks from [detail] in the foreground to [subject] behind it.
Use it for UGC, documentary, urgency
Handheld [shot size] of [subject], slight natural sway, at [height] height.
Use it for letting the action carry the shot
Locked-off [shot size] of [subject]; the camera does not move.
Use it for flat lays, food and desk scenes
Top-down shot of [subject] on [setting], slow [direction] drift, soft even light.
Use it for texture and small details
Macro slide across [detail] of [subject], extreme close-up, shallow focus.
Use it for a moment of shock or realization
Dolly zoom on [subject]: the camera moves in while the lens zooms out, [setting] stretching behind.
Use it for energetic tours of a space
Fast FPV fly-through of [setting], low and smooth, banking past [detail].
Most good prompts are 30 to 80 words: enough for the three layers, short enough that every word counts. A timeline adds a line per beat.
English is safest. Seedance 2.0 is also documented for Japanese, Spanish, Portuguese and Indonesian. In other languages the model may follow the prompt less closely. Keep references such as @image1 exactly as written.
No, the same structure works for both. Mini is the cheap place to test wording at 480p before you render the shot again on Fast at 720p.
Say who speaks and put the words in quotes, for example: she says: "It sinks in fast." Keep lines short. Audio is on by default and costs the same.
Only with permission. The pre-check flags real people and brands, and the model may refuse the job. A refused job is not charged.