ByteDance: Seedance 2.0 Fast

Seedance 2.0 Fast is ByteDance Seed’s speed-oriented edition of the Seedance 2.0 family. It supports joint video and audio generation and a multimodal brief built from text, images, video and sound. Its creative capabilities include new reference-led scenes, edits to existing footage and extensions of a scene, while its documented output levels are 480p and 720p.

Multimodal reference: assigning roles to the material

An image can supply a character, product, composition or visual treatment. A reference video can supply action, camera movement or an effect. Reference audio can supply voice timbre, musical melody or dialogue content. Explain which part of each source you want to carry into the new scene; a reference is more useful when it has a specific creative role.

For example: “Use the ceramic lamp in Image 1 as the product. Follow the slow camera orbit from Video 1, but place the lamp in a quiet reading corner. Use Audio 1 for the musical mood.” The intended output is a new scene combining the chosen appearance, motion and sonic direction, rather than an unspecified imitation of every detail in the references.

The documented combinations include images with audio, images with video, video with audio, and all three together. Text with audio alone, and audio-only input, are not supported. The model also supports text-to-video and image-led generation from a first frame or a first-and-last-frame pair. When an exact endpoint is central to the brief, use the dedicated endpoint workflow; describing a reference image as an endpoint in an omni-reference prompt is indirect control.

Editing an existing video

The model can add, remove or modify elements, replace a subject and redraw or repair parts of a scene. Identify the source video, the object to change and the elements that should stay. For an addition, describe where it appears and when. For a removal, identify the object consistently across the clip.

A practical edit brief might replace a plain bottle in a demonstration with the bottle from Image 1, while asking to preserve the hand movement, background and camera path of Video 1. Review both the replacement and the areas meant to remain stable. Clear preservation instructions make the intended edit easier to evaluate than a general request to “improve” the footage.

Extending and connecting clips

Seedance 2.0 Fast supports forward or backward extension and connecting two or three clips with generated transitions. For an extension, state whether the new action happens before or after the supplied video and describe that action. For a connection, give the clips in their intended order and explain the visual event that joins them.

A simple extension usually returns the newly developed continuation around the original footage rather than assuming the entire source clip is repeated. If the output should contain the original segment, say so in the brief. When filling a transition between two or three videos, the documented workflow can combine original and newly generated content. This is useful for linking a wide product view to a detail shot through a purposeful camera move.

Output and edition differences

The documented Fast output is MP4 at 24 fps, with 480p or 720p, 8-bit color depth and 4–15-second clips. Supported ratios are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. The standard Seedance 2.0 edition separately lists 1080p and 4K output; those levels are not part of the Fast specification. The 2.0 family supports returning the last generated frame, but does not support the separate Draft preview mode.

Audio can be generated with the footage or omitted for a silent clip. Generated audio is mono. 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 a spoken scene, name the speaker and put the line in quotation marks, then specify the delivery and relevant sound effects. Keep the required performance appropriate to the selected clip length.

Preparing references and observing input limits

The documented reference limits are 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 or audio clip must be 2–15 seconds long; the combined video duration is at most 15 seconds, and the combined audio duration is separately at most 15 seconds. Video references use MP4 or MOV, at 24–60 fps, and must not exceed 200 MB each. Audio references use WAV or MP3 and must not exceed 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; the total request-body limit is 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 assets and refer to them consistently as Image 1, Video 1 or Audio 1. Use the visible subject’s name or distinguishing feature as well as the asset number. Match endpoint-image framing to the intended video ratio, since a mismatch can crop the image or produce a jarring change in framing. Keep the brief focused on the features to transfer, the event to generate and what should remain unchanged; a long prompt can scatter the instructions and leave secondary details unresolved.

Shot planning, text and known behaviours

Organize a complex brief as Shot 1, Shot 2 and Shot 3, with a clear subject, action, position and sound for each. The guide recommends one camera movement per shot and says precise per-shot timestamps are unstable. Give each character a stable label and a few distinguishing features; use a separate clear headshot and full-body reference rather than a combined character-turnaround sheet when facial identity matters.

The model can generate common text, including slogans, subtitles and speech bubbles, with directions for its color, style, position and entrance. Common characters are preferred to rare characters or special symbols. Check the letters in motion. Unwanted subtitles, logos and duplicate characters are documented behaviours; explicit constraints and clean references can reduce them, but do not guarantee their absence. More than four reference people is reported as a stability risk, not a hard file-count limit.

Repeated extensions can compound image-quality loss, and an extension join can jump or roll back. Limit stacked continuations and inspect the join in the edit. Referenced voice timbre may drift, narration can end with a click, and Chinese polyphonic or uncommon characters may be mispronounced. Check the spoken line and ending audio; a short audio fade can help with a cutoff. These checks are part of preparing the creative result, especially when speech or character continuity carries the story.