Seedance 2.5 prompt format

How to write a prompt Seedance 2.5 reads the way you meant it, with references, lines and sounds placed where you intended. Two shapes: image-to-video (the first frame is the request's image_url, no references, no labels) and reference-to-video (every image is numbered in the prompt, the first frame is @Image1 only because the prompt says so, and a dialogue track can be @Audio1). The requests are in seedance-2-5-image-to-video and seedance-2-5-dialogue. Worked examples, one per mode, are in references/.

How to handle @ names, entrances, subjects not in the frame and reference style is shared by every video model: guide-animate-a-shot. Never put a person's name or a handle in the prompt outside {}: Seedance renders it as on-screen text.

Layout

Sections joined by blank lines: the header (reference-to-video only), the description, the footer.

Image 1 is the first frame: the video opens exactly on @Image1.
@Image2 and @Image3 are the woman already shown in @Image1, rendered in the visual style of @Image1.
@Image4 is the young man, rendered in the visual style of @Image1; not in @Image1 at the start, appears during the shot.
<audio lines, if a track is attached>

<description>

Sound effects: <rain tapping on the glass>, <footsteps on wet asphalt>.
Single continuous shot, no cuts. Do not generate duplicate versions of a character or a twin effect; keep only one instance of each character in each frame.
No BGM. No subtitles. No dialogue and no narration; everyone keeps their mouth naturally closed.

Image-to-video has no header, and its footer's second line is only Single continuous shot, no cuts. (the duplicate-character sentence is for references). With spoken words and no track, the last line is just No BGM. No subtitles.

Labels (reference-to-video)

• @Image1 is the first frame, always. Reference pictures follow from @Image2, in the order of image_urls. Up to 5 subjects and 29 reference pictures.
• A subject is cited by its first picture's label; its binding line names all of them ("@Image2 and @Image3 are the woman…", is for one picture, are for several).
• Binding line for a subject in the frame: <labels> is/are <noun> already shown in @Image1, rendered in the visual style of @Image1. For one who enters: <labels> is/are <noun>, rendered in the visual style of @Image1; not in @Image1 at the start, appears during the shot. Keep "rendered in the visual style of @Image1" as written: a reworded per-reference version did no better.
• The noun is a few plain words starting with "the" ("the young man", "the golden dog"), never a name or a label.
• In the description, cite each subject by its label at every mention; at first appearance follow it with the same noun ("@Image2, the woman,"). Never cite an @Image number that isn't bound.

The description

Under 200 words, one continuous shot, in order:

1. The style in explicit words, from the first frame: "Live-action, cinematic", "Hand-drawn 2D cartoon animation", "Stylized 3D CG animation". Seedance pulls a photo reference toward live action, so a non-live-action style must be named.
2. "The shot opens exactly on @Image1:" (image-to-video: "on the first frame:") then the setting, the light, and where each person or thing already in frame is and how the frame shows them. Carry clothing, colours and props through.
3. The action in the order given, ending on the result.

• Timing: integer-second marks only, 1 to 3 ("At the 1-second mark, @Image3, the young man, walks in from the left edge of the frame."), then "then" for everything after. A clip may be only 4 s long.
• Camera: exactly one movement for the whole shot (static, push in, pull out, pan, tilt, tracking, orbit, dive, crane up, or a rack focus said with what it does). Never two moves, never a cut.
• Big, clear, realistic motion. On-screen text in double quotes, verbatim.

Speech

• Exact words supplied: each line in its own {} right after its speaker and a short delivery, verbatim, in spoken order: @Image2, in a bright, breathless voice, says: {Morning! Did I miss the eight-fifteen?}. A subject speaker is cited by label, anyone else by noun phrase. Never a name before a line.
• No exact words: no { or } anywhere and no speaking verb for anyone (says, talks, explains, asks, mouths); write the visible gesture. Smiles and laughs closed-mouthed. The footer's "No dialogue and no narration; everyone keeps their mouth naturally closed." covers the rest.

Audio lines

With audio_urls attached (reference-to-video only), add to the header, after the binding lines, one of:

• Timed (preferred): the track runs the clip's full length with each line at its real time.

  @Audio1 is the finished recording of every spoken line, spoken on screen by the characters, never as narration or voiceover.
  From 0.09 to 1.48 seconds, @Image2 speaks {<the line>} from @Audio1, with the mouth of @Image2 moving in clean sync with it.
  

  one From … to … seconds line per spoken line, and replace the footer's last line with: No BGM. No subtitles. The only speech is the recorded dialogue (@Audio1), each line heard once, exactly as recorded; no other dialogue and no narration. Each character's mouth moves only during their own recorded line and stays closed otherwise. Without the timing, Seedance re-paces the line: it started up to 1.3 s late and ran about 10% long.
• Untimed window: @Audio1 is this shot's dialogue track: every line in {} is spoken with the voice, words and timing of @Audio1. The footer's last line is then just No BGM. No subtitles.

Sound effects

1 to 4 short sounds the scene and action make ("a ceramic mug set down on wood"), each in <> on the Sound effects: line. Never a voice, speech, singing or music, never something only visible. A quiet scene gets one or two. No @, {} or <> inside an item.

Check it

• Every subject's label in the description; no unbound @Image number; @Image1 only with references.
• Braces present exactly when there are spoken words.
• No @ handle or name outside {}.
• One camera move, marks only at 1 to 3 s.
• Footer lines exactly as above.