Video Model Seedance
This Skill is the Seedance model adapter layer. Before starting, it follows the unified model adapter contract of the video-generation dependency; this file only stores Seedance-specific differences. It compiles Seedance current model prompts from existing SHOT.md, reusable materials, and audio requirements, without redeciding shot creativity or long-term project facts.
1. Fixed Model / Entry Point / Task Mode
First confirm the Seedance version, entry point, and task type the user is actually using. Current focus is on adapting Seedance 2.5; if the user is using 2.0, 1.5, or a third-party Router, do not directly apply Seedance 2.5 parameters.
Common task modes:
- reference-based generation;
- standard video generation;
- storyboard / keyframe;
- video editing;
- extension;
- audio / voice reference.
Exact duration and reference material limits should be verified per the current references/MODEL-GUIDE.md.
2. Map Reference Materials First
Seedance reference generation first clearly describes each input:
- Which one is a character / subject;
- Which one is the scene;
- Which one handles product or prop structure;
- Which video only references motion / cinematography;
- Which audio handles sound or timbre.
Do not let a single reference video implicitly carry character identity, clothing, environment, and motion simultaneously. Only declare the parts you wish to transfer currently.
3. Structured Prompts
Complex reference-based shots can be organized as needed into:
- reference mapping;
- main idea;
- timeline;
- consistency;
- essential constraints.
Simple shots do not need to mechanically write all five sections. The structure is only used to reduce ambiguity, not to lengthen the prompt.
4. Timeline
Seedance is well-suited for using a clear timeline in longer generations. The timeline describes what is seen, who does what, how the camera moves, when dialogue/sound occurs, and how the next segment connects.
The number of actions must match the selected duration. Internal hard cuts should only be used if the current shot design inherently requires them.
5. Consistency
Only lock truly persistent content, such as character identity, product structure, scene, costume, cinematographic style, or sound. Information duplicated in the reference mapping should not be endlessly repeated in the prompt.
6. Editing / Extension
In video editing mode, explicitly use editing action words and only modify what the user specifies; content not requested to change should be kept as much as possible.
Extension tasks should prioritize maintaining input/output format and audiovisual continuity; precise format requirements should be verified against the current official entry point.
7. Return to video-generation
Output the Seedance model / entry point / task mode, a reference mapping written clearly in upload order, the current prompt, settings the user needs to confirm (duration / ratio / output format, etc.), and the explicit purpose of each reference.
One-shot generation packages and result re-integration are still handled by video-generation.
Completion Criteria
Each Seedance reference has a clear role; the prompt structure matches the task mode; the timeline does not exceed the actual duration; precise capabilities come from current official documentation, not model memory.