Reference video and audio

On a paid plan, the Seedance models take up to 2 reference videos (MP4 or MOV) and 1 audio track (MP3 or WAV) alongside the prompt and images. A reference video guides motion and content; an audio track gives the video its sound.

• Only the first 5 seconds of each reference video are sent, and that is what you are charged for. A shorter clip is charged for its own length.
• The price shown already includes the reference. Reference footage is billed as input, so a render with one costs more than the same render without.
• A reference clip shorter than 2 seconds is refused by the model.
• Attaching an audio track means the video has sound, whatever the Sound switch says.
• The audio track is sent whole, and the model takes 2 to 15 seconds of it on the 2.0 models and 2 to 30 seconds on Seedance 2.5. Outside that the render fails and is refunded; trim it first.
• On the 2.0 models an audio track needs a picture, a person or a reference video with it. Seedance 2.5 can take audio on its own.
• MiniMax H3, LTX-2.5 and Wan 2.6 do not take reference video or audio.

Telling the model what to take from them

• Videos and audio are numbered separately from pictures, in the order attached: Video 1, Video 2, Audio 1.
• Say what each is for: "the camera movement in Video 1", "the dance moves in Video 2", "the voice in Audio 1", "Audio 1 as the background music". When a reference shows something accurately, say to follow it rather than describing it again.
• Avoid editing words about a reference clip ("add", "replace", "remove", "change to"). They can make the model treat the request as an edit of that clip, which a new render cannot be, and it fails (refunded). Describe the video you want instead.
• A prompt that says "continue" or "extend" about a reference clip can make the model treat the request as a continuation of that clip. When that happens we render it once more in the clip's own shape, on the same charge, so it still arrives, but it leans on the clip and ignores the shape you picked. To carry a finished video on, use Extend.
• A reference with text on screen can bring text or subtitles into the result.
• Fewer, well-chosen references work better than the most you can attach: one or two pictures of a person, one of the place, one clip and one track is plenty.

Your own videos make good references, including for the next scene with the same characters — see Characters.