Seedance 2.5 Prompt Guide

ByteDance’s official prompt guide suggests treating Seedance 2.5 as a visual storyteller and writing the prompt in three parts. Think of it as a director’s brief: the idea in one line, the plot in order, then the rules for the whole shoot.

1 · A one-sentence summary

Subject, location, event, style and camera movement in a single line — the whole video at a glance.

2 · The plot, segment by segment

Split the video with timestamps (0–3s, 3–8s…) or shot numbers (Shot 1, Shot 2…). For each segment, describe what we see, how the camera moves, what happens, and any dialogue or sound.

3 · What stays the same

Finish with the details that must hold for the whole video: camera angle, setting, light, sound and atmosphere.

Put together, an 8-second clip looks like this:

A courier in a yellow jacket races across rain-soaked rooftops at night — handheld action, cinematic color grade, the camera chasing close behind.
0–4s: She sprints toward the edge of the roof, splashing through puddles, city lights blurred behind her.
4–8s: She leaps the gap to the next building, lands and rolls; the camera follows her over the edge.
Keep the yellow jacket, the night rain and the handheld camera throughout. Audio: heavy breathing, splashing footsteps, a distant siren, no BGM.

Write in positive terms — describe what should be on screen rather than what shouldn’t. The exceptions are subtitles and audio, where “no subtitles”, “no BGM” and “no audio” work reliably.

A single Seedance 2.5 generation can run up to 30 seconds — long enough for a small scene with a beginning, middle and end. Unlike Seedance 2.0, which only followed shot numbers, Seedance 2.5 responds to timestamps, using whole seconds as the unit. Three ways to control timing:

• Time ranges — 0–3s … 3–7s … 7–15s or [1s–4s] … [4s–8s]. Keep them continuous, with no gaps.
• Time points — “at the 5-second mark, a flash of lightning fills the frame”.
• Relative timing — “he stands frozen; after 3 seconds, the crowd turns to look”.

Balance the content against the time. Too little in a range and the model improvises; too much and you may get extra cuts or skipped beats. Timestamps are for story beats, not rapid-fire motion — don’t ask for something like “shakes his head three times per second”.

[0–5s] Aerial shot gliding low over dark Atlantic waves at dawn, mist hanging over the water.
[5–10s] The camera rises and tilts up to reveal a lone lighthouse on a cliff, its beam sweeping through the fog.
[10–15s] Slow push in toward the lighthouse as the first sunlight breaks through the clouds and lights the cliff edge. Cinematic color grade.
Audio: wind and crashing surf, distant gulls, no music.

Standard film terms can be written directly. For a niche technique, add a short explanation after the term — for example, “rack focus: the foreground trees blur as the character behind them comes into focus.”

Shot sizes and angles work the same way: extreme wide, wide, medium, close-up, plus low angle, overhead and first-person. For a transition, say when it happens and how: “at the 5-second mark, a quick wipe to the left.”

Seedance 2.5 creates the soundtrack together with the video, so the prompt doubles as your sound brief. It can generate speech natively in more than 10 languages. On Moluri audio is on by default; switch off “Generate audio” in the settings for a silent clip.

• Dialogue: say who speaks and in what tone — and the language or accent if it matters — then put the line in quotes: She says quietly in American English, nervous: “I’ve never seen him before.”
• Sound effects and ambience: describe them plainly — “rain on a tin roof”, “a kettle ticking as it cools”, “muffled chatter”.
• Music: describe the score you want, or write “no BGM” to keep only natural and action sounds.
• Subtitles: add “no subtitles” if you don’t want text burned into the frame.

First/last frame. Your image becomes the opening frame, and the video’s aspect ratio is locked to it. If you add a last frame, use the same aspect ratio — a last frame with a different ratio gets stretched. The image already shows the scene, so spend the prompt on what moves and changes: the camera, the action and the sound.

Reference video. Upload up to 9 images and refer to them as @image1, @image2 and so on, in upload order. Give each image one job, bind it in its own phrase, and say what to take from it:

@image1 is the main character. @image2 is the setting. The woman from @image1 walks through the night market from @image2, wearing the leather jacket from @image3. Match the warm lantern lighting of @image2. Handheld tracking shot. Audio: market chatter, sizzling food stalls.

• Order matters. With several characters, upload their images in the order they first appear — otherwise characters can get swapped.
• Map in the prompt, not on the image. Writing a name on a character’s picture isn’t enough; say “the knight in @image1” in the text.
• Don’t re-describe what the image shows. If a reference already nails the look, just point to it.

Copy a template, replace the brackets, and delete anything you don’t need.

[Shot size and camera movement]. [Subject] [one clear action] in [setting], [light and time of day]. [Style or look]. Audio: [sound effects and ambience], [music or “no music”].

[One line that sets up the scene: place, light, style and who is in it.]
[0–Xs] [Camera]. [What happens.]
[X–Ys] [Camera]. [What happens next.]
[Y–Zs] [Camera]. [How the shot ends.]
Audio: [ambience and effects throughout], [music or “no music”].

[Shot and setting]. [Character A, short description] and [Character B, short description].
[0–Xs] [Action]. [Character A] says in [language and accent], [tone]: “[line]”
[X–Ys] [Reaction or reply].
Audio: [room tone], [music or “no music”].

@image1 is [the main character / the product]. @image2 is the setting. [The character from @image1] [action] in [the place from @image2], wearing [the jacket from @image3]. [Camera]. [Light and style]. Audio: [sounds].