Seedance 2.0 Omni-Reference — Multimodal Input Guide

Text: natural-language prompt. Images: often up to 9 (e.g. 30 MB each in some docs). Video: up to 3 clips, often 2–15 s total, ~50 MB per clip. Audio: up to 3 files, often ≤15 s total, ~15 MB each. Total of up to 12 reference files in one go. The model uses them for layout, motion, camera, style, and sound as directed by your prompt and @ tags.

You can refer to uploaded assets in the prompt with @ tags (e.g. @Image1, @Video1, @Audio1). Examples from public docs: “@Image1 as the first frame,” “Reference @Video1 for camera movement,” “Use @Audio1 for background music.” This gives precise control over which image drives character, which video drives motion, and which audio drives music or dialogue.

Common @ tag patterns for Omni-Reference: (1) First-frame lock: '@Image1 as the opening frame, character walks toward camera' — pins the starting composition. (2) Character consistency: 'Same character as @Image1, wearing the same outfit as @Image2' — locks identity across shots. (3) Camera replication: 'Replicate the camera movement from @Video1, apply to new scene with @Image1 as subject' — transfers motion path. (4) Audio-driven: 'Use @Audio1 as background music, lip-sync dialogue to @Audio2' — separates music and voice tracks. (5) Multi-reference combo: '@Image1 as character, @Image2 as background, reference @Video1 for camera motion, @Audio1 for ambient sound' — full scene assembly using 4 references. Always explicitly state each asset's role in the prompt; unnamed references may be ignored or misinterpreted.

Seedance 2.0 generates video and audio in a single joint process (not post-dubbing). It supports stereo output, lip-sync (including multiple languages in public reports), and alignment of music and sound effects with the picture. Useful for ads, MV, and dialogue-heavy clips.