ByteDance: Seedance 2.0

Seedance 2.0 generates video from a written scene or a combination of image, video and audio references. The exact selector doubao-seedance-2-0-260128 identifies the standard 2.0 profile. Its documented workflows include new video generation, editing and extension of existing footage. Use references to communicate appearance, movement, camera language and sound, then state which parts of the scene they should control.

Text can be supplied alone. The documented multimodal combinations also allow images or video with optional text, images with audio, video with audio, and images, video and audio together. Audio therefore accompanies visual media in these listed combinations; it is not described as a standalone audio-to-video request. Write the intended result clearly even where text is optional, so the role of each reference is explicit rather than left to be inferred from unrelated assets.

The content array distinguishes text, image_url, video_url and audio_url items. Multimodal references use roles such as reference_image, reference_video and reference_audio. An image can define the person, object, location or look. A video can demonstrate a performance, movement or camera view, and audio can establish a track or timing. State the assignment in the brief: keep the subject from the image, follow the camera view of the video and use the audio as the background. Do not ask every reference to define every property at once.

The documented advertising example combines two product images with a video and an audio reference. It describes an opening view, an action sequence, close product detail and a closing composition while keeping the background voice consistent. The same structure can organize a short showcase: establish the object, demonstrate one action and finish on its recognizable feature. For an edit, identify the source footage and the element that should change; for an extension, explain the action that continues beyond the existing scene while preserving the subject and movement direction.

The 2.0 profile documents whole-second duration from four to fifteen, or -1 for an automatically selected duration within that range. The default is five seconds. Its resolution options are 480p and 720p, with 720p as the default; 1080p is excluded for 2.0. Fixed aspect ratios include 16:9, 4:3, 1:1, 3:4, 9:16 and 21:9. Adaptive ratio can choose framing from the prompt or reference: first-frame work follows the opening image, and multimodal work uses the intended scene or the first relevant media, with video taking priority over images. Check the task result for the actual ratio and duration.

Generated sound is controlled separately from reference media. generate_audio=true requests synchronized voices, effects and background music based on the text and visual content; false requests a silent video. The documented default is true, and quoted dialogue is recommended for speech. Specify who speaks and how the sound relates to the visible action, then inspect the actual wording, voice assignment and timing. Supplying an audio reference and enabling generated sound are distinct request choices.

The 2.0 guide excludes the older frames and camera_fixed controls and does not support flex offline processing. Use the duration field and describe the camera movement in the scene itself. Seed and watermark remain documented controls; the guide says the same request and seed can produce similar results without guaranteeing exact identity. Avoid treating a seed as proof of reproducibility, or a repeated request as an already inspected clip.

Video-token usage is calculated as (input-video duration + output-video duration) × output width × output height × output frame rate / 1024. Requests containing input video and those without input video have distinct billing categories. The task query reports usage.completion_tokens and usage.total_tokens; retain these reported quantities with the completed clip.

The workflow is asynchronous: submission returns a task identifier, and the completed task supplies content.video_url together with reported duration, resolution, ratio, frame rate and token usage. Preserve the identifier and save the completed video with the original references and prompt. The expiry field defaults to 172800 seconds and accepts 3600–259200 seconds from task creation; a task that exceeds the interval becomes expired. Review the entire output, including the edit or continuation join, for character appearance, object detail, spatial relationships, camera continuity and sound-to-picture timing.