seedance-20
Seedance 2.0 quad-modal AI filmmaking skill for text-to-video, image-to-video, video-to-video, and reference-to-video workflows.
Operating Principle
Direct the model. Do not micro-manage it. State the intended subject, action, feeling, camera behavior, lighting logic, reference roles, audio layer, and constraints. Use references to show what should be preserved; use text to say what should change.
Workflow Router
- Vague idea, concept only, or no clear scene: load
[skill:seedance-interview]. - Clear concept or reference assets supplied: load
[skill:seedance-prompt]. - Need short Chinese prompt or prompt compression: load
[skill:seedance-prompt-short]and the relevant vocabulary skill. - Image-to-video: load
[ref:i2v-guide]and describe only what the image cannot show. - Video/reference-to-video: load
[ref:reference-workflow]; map each reference to a role. - Camera, motion, lighting, character, style, VFX, or audio specialization: load only the matching sub-skill.
- Bad output: load
[skill:seedance-troubleshoot]; then load the specific failure-area skill. - Copyright, celebrity, brand, franchise, real-person, or likeness risk: load
[skill:seedance-copyright]before finalizing. - Blocked or degraded prompt: load
[skill:seedance-filter]and preserve intent while changing risky surface wording. - Platform/API/integration questions: load
[skill:seedance-pipeline]plus[ref:api-status].
Current Platform Status Rule
Platform status changes quickly. For API availability, upload limits, face/portrait authorization, pricing, and regional access, do not rely on static memory. Load [ref:api-status] and check its last_verified date. As of 2026-04-27, official ByteDance/BytePlus sources describe Seedance 2.0 as supporting text plus image, audio, and video references, and BytePlus ModelArk publishes video generation API documentation. Real-person likeness workflows remain authorization-dependent and surface-specific.
Prompt Construction Rules
- Preserve reference tags before adjectives:
[Image1]identity,[Video1]motion,[Audio1]rhythm or voice. - Use one primary camera move per short clip unless the user requests a multi-shot sequence.
- For multi-character scenes, assign actions to named character tags. Avoid ambiguous pronouns.
- For I2V, do not re-describe visible static details. Add motion, camera, timing, transformation, lighting change, and audio.
- Remove hollow quality boosters. Prefer physical, observable, production-specific language.
- Rewrite protected IP, celebrity, brand, and real-person content into original archetypes unless the user supplies a clearly authorized workflow.
References
Load references selectively: [ref:api-status], [ref:platform-constraints], [ref:json-schema], [ref:prompt-examples], [ref:quick-ref], [ref:storytelling-framework], [ref:genre-guides], [ref:reference-workflow], [ref:i2v-guide], [ref:intent-vs-precision], [ref:source-registry].