ByteDance: Seedance 2.0

Seedance 2.0 is ByteDance Seed’s standard model for multimodal audio-video creation. It combines text, images, videos and audio in a joint generation architecture, with separate creative workflows for reference-led scenes, video editing and video extension. It is useful when the brief needs both a coherent performance and specific visual or sonic information from prepared materials.

Multimodal reference and directorial control

Images can define the subject’s appearance, the visual style and the composition. Video references can guide action, camera movement and effects. Audio can guide voice timbre, music or dialogue. Define the role of each material, then describe the scene that should result. The manufacturer emphasizes control over performance, lighting, shadows and camera language.

For a product story, prepare a product picture, a reference for the camera move and an audio cue. A brief might be: “Preserve the blue glass vase from Image 1. Use the slow sideways tracking movement from Video 1. The vase sits on a sunlit shelf, with soft instrumental music based on Audio 1.” The intended output is a newly staged scene that combines selected reference information.

Reference inputs can combine image and audio, image and video, video and audio, or all three. Audio-only input and text-plus-audio without image or video are not supported. Traditional text-to-video, first-frame generation and first-and-last-frame generation are also documented. If fixed opening and closing compositions are important, use the dedicated endpoint workflow: naming a reference image as an endpoint inside an omni-reference prompt is only an indirect way to guide it.

Editing footage

The model can replace a subject, add or remove objects, change elements and redraw or repair parts of a video. Describe the specific change and identify the source footage. For an addition, include appearance time and position; for a removal or replacement, state the object precisely and name what should remain.

For example: “Replace the desk lamp in Video 1 with the brass lamp in Image 1. Keep the camera movement, desk arrangement and lighting direction.” The desired result is an edited version of the scene with the intended substitution. Inspect the unchanged areas as well as the new object, particularly through occlusion, movement and shot changes.

Extending and connecting a sequence

Seedance 2.0 can extend a video forward or backward, and connect up to three supplied clips with generated transitions. Specify the temporal direction and the new event. For a bridge, order the clips and describe the transition between each pair, such as a camera entering a doorway in one shot and arriving in the next environment.

A straightforward extension typically returns the continuation developed around the source footage; include a clear instruction if the final result should also repeat the original segment. When completing a connection between two or three clips, the documented workflow can include both original material and new transition content. This is useful for developing an opening or ending, or connecting prepared scenes without leaving the joining action unspecified.

Standard-edition output and 4K

The standard edition documents MP4 at 24 fps and durations from 4 to 15 seconds. The output levels are 480p, 720p and 1080p with 8-bit color depth, plus 4K with 10-bit color depth. The supported ratios are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Adaptive framing and automatic duration selection are also documented. The Fast edition has its own narrower 480p/720p output specification.

The 4K output uses H.265/HEVC encoding. Ten-bit encoding preserves additional tonal gradations and smoother color transitions, which matters for material with subtle light or color gradients. Check that the intended playback and editing environment can decode that format; some browsers and devices do not play it directly. A compatible container does not by itself ensure compatible video decoding.

The model can generate synchronized sound or a silent video, and it can return the last generated frame for a subsequent scene. Generated audio is mono. The 2.0 family does not support the separate two-stage Draft preview mode, so that preview feature should not be inferred from another Seedance edition.

Reference preparation and substantive limits

The documented reference capacity is up to nine images, three videos and three audio clips. Direct reference-image or video uploads containing real human faces are not accepted in this documented producer profile; portrait work uses its supported portrait-asset workflow. Each video and audio clip must be 2–15 seconds long. All video references together must fit within 15 seconds, and all audio references together have their own 15-second total. Video files use MP4 or MOV at 24–60 fps, up to 200 MB each; audio files use WAV or MP3, up to 15 MB each. The listed image formats are JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC and HEIF. Images must be smaller than 30 MB each, and the request body must not exceed 64 MB. In the documented reference-input profile, image width and height must each be strictly between 300 and 6,000 pixels, with a width-to-height ratio strictly between 0.4 and 2.5. Reference videos allow 300–6,000 pixels in each dimension and ratios of 0.4–2.5, inclusive, with 407,696–8,295,044 total pixels. MP4 or MOV videos use H.264/AVC or H.265/HEVC video and AAC or MP3 audio.

Order the files deliberately and refer to each by its media type and number: Image 1, Video 1 and Audio 1. The number follows the order among files of that type. Explain which traits to transfer and which source content should not define the new scene. Documented prompt languages include English, Spanish, Indonesian, Portuguese and Japanese. Keep the spoken language consistent within a dialogue; the guide advises against mixing languages except for proper names. For voice-guided work, separate the speaker, the exact line and the delivery direction.

Prepare first and last images with compatible framing. A mismatch between the input picture and the output dimensions can cause a crop or an abrupt visual change; matched dimensions and adaptive framing are documented remedies. Keep the description concentrated on the event and the reference roles. The manufacturer warns that long prompts can scatter information, causing details to be ignored. Review the resulting motion, reference continuity, audio performance and joins between scenes as distinct parts of the creative result.

Shot planning and text direction

For a complex scene, organize the brief as Shot 1, Shot 2 and Shot 3, specifying the camera, subject action, position and sound in order. The guide recommends one camera movement per shot and prioritizing slow, connected actions. Precise per-shot timestamps are unstable, so an event sequence is a better control than forcing several exact time windows.

The model also generates common text, including slogans, subtitles and speech bubbles. You can direct the text’s color, visual treatment, location and entrance. Use common characters rather than rare characters or special symbols, and review the wording as it moves. For subject references, clearly bind each asset to the intended character or object. Product viewpoints can supply different visual information; character-turnaround collages can confuse facial identity. A separate clear headshot and full-body image is the documented character-reference recommendation.

Behaviours to inspect before an edit

The guide documents unwanted subtitles, logos, duplicate characters and style drift. Clean source material, consistent labels and explicit constraints reduce those risks, but do not guarantee that unwanted text or duplicate figures disappear. More than four reference people is reported as a stability risk, distinct from the number of files the model accepts.

Video extension can create a jump or rollback at the join, and repeated continuations can compound image-quality degradation. Inspect the boundary frames and limit stacked extensions. A referenced voice can differ from the source timbre; narration may end with a click or abrupt cutoff, for which an audio fade is a documented remedy. Chinese polyphonic, uncommon or similarly shaped characters can be pronounced incorrectly. For a precise animated effect, such as a countdown, a motion reference is more informative than a verbal request alone. These are concrete reasons to check identity, text, motion and sound separately when selecting the clip.