seedance-motion · Intent-First Choreography (v5.0)

This skill covers motion control, action choreography, and video extension for Seedance 2.0, prioritizing intent-driven descriptions over micro-management.

The Guiding Philosophy

For action, describe the intent and consequence, not the precise timestamps. Let the AI director handle the interpolation.

1. The Recommended Workflow: Intent + @Video Reference

This is the most reliable method for all action and fight scenes.

Step 1: Find a Reference Video

Find a real-world video clip (e.g., from a movie, a stunt performance, or a video game) that captures the style of action you want. Upload it as @Video1.

Step 2: Write an Intent-Based Prompt

Describe the high-level action in 1-3 sentences. Use degree adverbs and physics consequences. Then, explicitly tell the model to reference the uploaded video.

// Recommended Prompt Structure

Characters: A references @Image1; B references @Image2.

Choreography: The archer fires two arrows; the mage deflects them with a violet energy shield, then closes distance and blasts the archer into a tree with a shockwave. The archer draws a short blade and counter-attacks in close combat.

Reference: Reference the fight actions, character movements, and camera work from @Video1.

Style: Match the gritty, handheld style of @Video1.

Why this works: The @Video1 reference provides the model with a rich, dense, and unambiguous understanding of the desired motion, physics, and camera language, which consistently outperforms any text-only description.

For more on the @reference system, see [ref:reference-workflow].

2. The Text-Only Workflow: Intent-Based Description

Use this when you don't have a reference video. The key is to keep it simple and enforce the "One Action Per Shot" rule.

// Text-Only Fight Scene Example

Characters: A references @Image1; B references @Image2.

Shot 1: A throws a right hook at B's jaw.
Shot 2: B ducks under the punch and sweeps A's legs.
Shot 3: A jumps, landing a spinning back kick to B's shoulder.
Shot 4: B staggers backward two steps, recovering his balance.

Camera: Medium shot, tracking the action. Slight handheld shake on impacts.
Physics: Dust puffs up from the ground on the leg sweep. A wet impact sound accompanies each hit.

Key Principles for Text-Only Action

One Action Per Shot: Do not chain multiple distinct actions (e.g., punch, block, kick) into a single sentence or shot. Break them down.
Degree Adverbs: Use words like violently, gracefully, slowly, frantically to guide the model's interpretation of the action.
Physics Consequences: Describe the results of the action. Dust erupts, sparks fly, water sprays, the character staggers.

3. Experimental Workflow: Micro-Choreography

⚠️ Warning: This is an advanced, experimental technique that is unreliable for most users and often results in jitter, morphing, and failed generations. Use the Intent-Based workflows above for production.

Micro-choreography involves specifying actions with timestamps or in a grid format. While it offers the highest potential for control, it is also the most likely to fail.

The Grid Method (25宫格)

Timestamp Method

0-1s: A throws a punch. 1-2s: B blocks. 2-3s: A follows with a kick.

When to use: Only for short, highly technical sequences where the exact timing of each beat is critical and you are prepared to iterate many times to get a usable result.

Diagnostic Tools

Use these concepts to diagnose failing motion prompts, not as prescriptive rules for building them.

Beat Density: If your output is blurry or jittery, you may have too many actions packed into a short duration. The model can typically handle 1-2 distinct beats every 5 seconds. High-density prompts require the experimental micro-choreography format.
Timing Language: Use relative terms (eases in over 2 seconds) or descriptive adverbs (accelerates into a run) instead of hard timestamps for smoother, more natural motion.

Maintained by Emily (@iamemily2050)