Seedance 2.5 Flash is a faster edition of Seedance 2.5, so the prompt habits that work on Seedance 2.5 are the place to start. This guide collects the rules we use every day on thousands of Seedance videos, with examples you can copy.
Good news if you already write for Seedance 2.5: we expect your prompts to carry straight over to Flash. We'll confirm that with side-by-side tests on launch day and note any differences here.
The shape of a good Seedance prompt
Write like a director, in things the camera can see. Cover these, and skip any the idea doesn't need:
- Subject: who or what is on screen, in concrete terms.
- Action: what happens, and what is visibly different at the end.
- Scene: where, when, the light and the weather.
- Camera: one move per shot, like push in, pan, tracking, orbit or handheld.
- Sound: music, sound effects and any spoken lines.
Turn feelings into what we can see: "her smile fades and she sets the cup down" works better than "she feels betrayed".
Keep settings out of the prompt
Aspect ratio, length, resolution and sound on or off are settings, not prompt words. Don't write "16:9", "1080p" or "a 10-second video" in the text. Use the length only to judge how much can happen: a clip of 5 seconds or less holds one moment, or two shots at most.
Timing: stages first, seconds only when you need them
Seedance 2.5 understands whole-second ranges like 0-3s: and 3-7s:, but you rarely need them. By default, write the action as stages and let the model pace it. Use second ranges only when a beat must land at a set moment, back to back, about three seconds per beat, and never with decimals.
0-3s: Overhead shot of a chessboard mid-game; a young man's hand slides a white knight forward. <a soft wooden click>
3-6s: Cut to a close-up of his opponent, an older woman; her eyebrows rise and she leans back.
6-9s: Medium shot of both players; she tips her king over and offers a handshake. The older woman (amused, in English): {Well played.} No subtitles.
Dialogue, music and sound effects
Seedance makes sound with the picture, and it reads three kinds of brackets:
- Music in round brackets:
(soft piano fades in) - Sound effects in angle brackets:
<rain on a tin roof> - Spoken lines in curly brackets, with the speaker, delivery and language first:
Mia (quietly, in English): {I kept your letter.}
When someone speaks and you don't want captions, add "No subtitles." When nobody should speak, write "Dialogue: None." Plan about two to three words per second of a speaker's shot.
Using reference images, videos and audio
Seedance 2.5 can take many references at once (on Lumeta: up to 30 images, 10 videos and 10 audio clips; Flash takes up to 10 images, 5 videos and 5 audio clips). Give each one its own line at the start, saying what it supplies and what it must not bring along:
@Image1 (Mia): her face, hair and green raincoat; do not use the background.
@Image2 (the bakery): the shop front and morning light; do not use any people in it.
@Audio1: Mia's voice, its warmth and pace only.
A video reference supplies motion, camera or pacing, not the people in it, unless you say so. Write "follow the motion of @Video1", never "replace the dancer in @Video1": edit words can turn your generation into an edit of that video.
Words on screen
Put the exact words in double quotes and say where they appear: the words "CHILL BREW" fade in at the center in bold white letters. Keep it to a few common words. Printing a name on the product itself often reads better than an overlay.
A full example you can copy
A rainy night street in a neon-lit city, cinematic and moody, wet asphalt reflecting pink and blue signs.
A courier in a yellow rain jacket pedals a bicycle through puddles toward the camera, slow push in. <rain hammering on awnings>
Over her shoulder, a small dog shivers in a doorway ahead. She brakes hard, water spraying from the front wheel.
She crouches, unzips her jacket and tucks the dog inside. The courier (softly, in English): {You're coming with me.}
(Low, warm synth music fades in.) No subtitles.
What to avoid
- Names of famous people or characters. Describe the look instead; names can get a clip refused.
- Long lists of "no" words. There is no negative prompt. Short exclusions work for subtitles, music, duplicates and watermarks; for everything else, describe what you want.
- Cramming. If too much happens for the length, cut events instead of rushing them.
- Quality filler like "8K, masterpiece". Seedance 2.5 doesn't need it.
Test your prompts on Seedance 2.5 today
Lumeta is the cheapest place to run Seedance 2.5, and Flash joins it on launch day. Your prompts will be ready.
Prompt questions
Do Seedance 2.5 prompts work on Seedance 2.5 Flash?
We expect so: Flash is built on Seedance 2.5. We'll confirm it with side-by-side tests on launch day.
How long should a Seedance prompt be?
As long as the idea needs. A single moment often fits in 40 to 80 words, a multi-shot scene in 150 to 400. Detail starts getting dropped well before ByteDance's limit of about 1,000 English words.
What languages does Seedance 2.5 understand?
English works best for the prompt. Seedance 2.5 also reads Chinese, Spanish, Indonesian, Portuguese, Japanese, Malay, Thai, Arabic, Vietnamese and Korean, and spoken lines can be in the language you want.
Is the prompt different for Seedance 2.0?
Mostly the same, with one big difference: Seedance 2.0 ignores second ranges like "0-3s". Write numbered shots instead (Shot 1, Shot 2), about one shot per three to five seconds.
Built from ByteDance's own Seedance prompt guidance and our daily use of Seedance 2.5 and 2.0 on Lumeta. Flash-specific notes will be added after launch-day testing.