ByteDance: Seedance 2.0 Mini

Seedance 2.0 Mini is a separately named model in ByteDance’s Seedance 2.0 audiovisual family. Its manufacturer documents text-to-video, first-frame and first-and-last-frame image-to-video, multimodal reference generation, video editing and video extension. Sound can be generated together with the picture. Choose a creation mode according to the material you already have: a written brief, a fixed opening image, two boundary frames or a clip that needs an edit or continuation.

Mini has its own output profile. The manufacturer lists 480p and 720p video with 8-bit colour, 24 frames per second and MP4 output. Its supported aspect ratios are 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16; documented generation duration is 4–15 seconds. The producer’s duration control also allows automatic selection within that range. The full Seedance 2.0 model’s 1080p and 4K options are separate; they are not Mini settings. The family capability table lists returning the last generated frame, while Draft mode is unsupported.

Omni reference creates a new scene using selected information from the supplied material. An image can define a subject’s appearance, composition or visual treatment. A video can establish action, camera movement or an effect; an audio reference can establish voice timbre, melody or spoken content. Tell the model what to take from each reference instead of asking it to copy every property. For example, use a product photograph for shape and colour, a short camera-movement clip for the approach, and a sound reference for the atmosphere.

The family workflow allows up to nine reference images, three video clips and three audio clips. Video references are 2–15 seconds each, with no more than 15 seconds in total; audio references use the same per-clip and total duration limits. Audio alone and text plus audio without a visual reference are not supported reference combinations. Text-only creation is a separate supported mode. For a sequence, provide assets in the intended order and refer to them consistently as Image 1, Video 1 or Audio 1, using the order within each media type. Do not substitute an asset identifier for a numbered reference in the written prompt.

A first-frame image defines the opening of an image-led shot. The first-and-last-frame mode defines both boundaries. If those exact boundary images matter, select that dedicated mode rather than describing references as an opening and ending inside an omni-reference prompt. The latter is a prompt-directed effect and is not equivalent to a strict boundary-frame choice. Prepare the images for the output composition, and state the action that connects the starting and ending situations.

Video editing changes an existing clip instead of using it only as inspiration. Identify the video to edit, the object or region to change, its appearance time and position, and the details to preserve. A useful brief is “In Video 1, replace the mug with the vessel shown in Image 1; preserve the hand movement, framing, lighting and table arrangement.” For removal, name the unwanted object and the surrounding elements that must remain. The maker advises referring directly to the target video for editing or extension; wording it as a reference can instead communicate a request for a newly generated scene.

Video extension adds events before or after an existing clip. Describe the direction of continuation and what happens next or previously. The documented workflow can also connect two or three clips with generated transitions; the three input videos together still fit within the 15-second reference budget. The tutorial distinguishes a continuation that mainly includes the original clip’s boundary footage from a request that explicitly keeps the original content. State the desired composition of the result: only the continuation, or a scene containing the supplied clip and the new action.

For clear direction, separate the stable scene description from the sequence of events. Define each subject with two or three distinguishing visual features and a consistent label. Then describe actions, setting, lighting, camera movement, visual style and constraints. In a dialogue, associate each line with its speaker and relevant action. Use ordered shot labels rather than treating exact timestamps as reliable controls: the maker notes that precise timing can be unstable. A staged product brief might read: “Image 1 defines Product A. Shot 1: slowly approach Product A on the counter. Shot 2: a hand rotates it toward the light. Shot 3: settle on the front label. Soft room ambience; preserve the printed lettering.”

Prepare references for the documented family input rules. Supported image formats include JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC and HEIF; each image is smaller than 30 MB. The guide specifies image width and height strictly between 300 and 6,000 pixels and an aspect ratio strictly between 0.4 and 2.5. Video references use MP4 or MOV with H.264 or H.265 video and AAC or MP3 audio, are at most 200 MB each and have 24–60 frames per second. Their width and height are 300–6,000 pixels, aspect ratio is 0.4–2.5, and total pixel count is 407,696–8,295,044. Audio references use WAV or MP3 and are at most 15 MB each. These are the manufacturer’s input-preparation specifications, distinct from a particular application’s upload controls.

The maker’s portrait workflow requires specific preparation for images or videos containing real human faces: direct reference uploads are not the standard supported path. It documents trusted model outputs, preset digital characters and an authorized real-person asset workflow as separate solutions. Voice matching, pronunciation and the sound at a clip boundary also need inspection. The prompt guide recommends describing the intended voice characteristics and keeping dialogue style close to the audio reference. Review the final artifact for subject consistency, exact spoken words, action timing, frame transitions and ending audio; refine the particular element that differs from the brief.