Native audio and lip sync in Seedance

What native audio means

In older pipelines you generated a silent clip, then added sound with a separate tool and hoped the timing matched. A model with native audio produces the soundtrack as part of the same generation, so footsteps land when feet hit the ground and a character's mouth moves with the words. The trade-off is that you control sound through the same prompt as the picture, which takes a different writing habit.

Audio is a switch in the workspace. On the Seedance 2.x models it does not change the credit estimate; on Seedance 1.5 Pro it doubles the rate, so leave it off for silent 1.5 Pro drafts.

Three kinds of sound you can ask for

1. 01

   Ambience

   The background of the scene: rain, traffic, a humming fridge, wind through grass. Name two or three sources and their loudness. Ambience is the most reliable category.

2. 02

   Effects tied to action

   Sounds caused by what happens on screen: a door closing, a glass set down, footsteps on gravel. Describe the action clearly and the model will usually sync the sound to it.

3. 03

   Dialogue and voice

   Spoken lines with lip movement. Put the exact words in quotation marks, say who speaks, and describe tone and pace. Keep lines short; long monologues drift out of sync.

Music is the least controllable category. You can ask for 'a slow piano motif' and often get something suitable, but you cannot dictate melody, key or tempo, and you should not expect a licensed-sounding track. If music matters, generate the clip without music and score it in an editor.

Writing a dialogue prompt

Medium close-up of a woman in her forties, grey cardigan, standing in a small bookshop. She looks at the camera and says, warmly and unhurried: "We open at nine. The good chairs are upstairs." Quiet shop ambience, pages turning somewhere, no music. Static camera, soft daylight from the left.

• One speaker per clip gives the best sync. Two speakers work but require clear turn-taking in the prompt.
• Front-facing or three-quarter framing shows the mouth; profile shots hide sync problems but also hide sync successes.
• State 'no music' explicitly when you want clean dialogue; otherwise the model may add a bed.
• Numbers, names and unusual words are more likely to be mispronounced; spell them the way they sound if needed.

Ambience and effects prompts

A night market alley after rain. Neon signs reflect in puddles, a vendor flips skewers over coals. Sound: sizzling and crackling from the grill close to camera, distant crowd murmur, occasional scooter passing, water dripping from awnings. No dialogue, no music. Slow lateral tracking shot.

A carpenter sets a chisel on a workbench, picks up a wooden mallet and strikes a joint three times, then blows sawdust off the surface. Each strike is a dry, solid knock; the sawdust makes a soft puff. Workshop ambience, a radio faintly in another room. Close-up, static camera, warm tungsten light.

Reference audio on Seedance 2.0 and 2.5

Seedance 2.0 and 2.5 accept audio files as references alongside images and video. On Seedancee, 2.5 takes up to 10 clips and 2.0 up to 3; the provider caps individual clips and total duration, and the workspace shows the limits before upload. Two uses work well:

• Voice reference: a clean recording of a voice so generated speech matches its timbre and cadence. Use a single speaker, no background noise, and tell the prompt that the character speaks in the referenced voice.
• Rhythm reference: a beat or a short music clip so on-screen motion follows the tempo, for example cuts of movement landing on the downbeat. Say explicitly that the motion should follow the audio.

Where native audio still fails

• Long lines: past roughly ten words, sync tends to drift; split dialogue across shots.
• Multiple simultaneous voices: crowds are fine, but two people talking over each other is not.
• Specific music: you get a mood, not a composition.
• Silence: the model sometimes fills quiet moments; ask for 'near-silent room tone' rather than 'no sound'.
• Language: our tests are English-first; we will publish results for other languages in the Lab as we run them.

The Seedance Lab includes a native-audio test with a fixed spoken line and ambience cue, scored for line accuracy, lip-sync offset and ambience presence. Results are published per model with dates as they are run.