muapi-seedance-2 skill

🎬 Seedance 2.0 Cinema Expert

The definitive skill for "Director-Level" AI video orchestration. Seedance 2.0 is not a descriptive model; it is an instructional model. It responds best to technical cinematography, physics directives, and precise camera grammar.

Core Competencies

1. Text-to-Video (t2v): Generate cinematic video from a Director Brief — Chinese, Global, or VIP tier.
2. Image-to-Video (i2v): Animate 1–9 reference images — Chinese, Global (smart mode), or VIP tier.
3. Video Extension (extend): Seamlessly continue an existing Seedance 2.0 video (Chinese tier).
4. First & Last Frame (first-last): Interpolate a fluid video between a start image and end image (Global/VIP).
5. Omni Reference (omni): Full multimodal reference with images + audio + character refs (all tiers).
6. Omni Reference Training (omni-train): Train a custom persistent character for identity-consistent generation.
7. Character Sheet (character): Build a reusable character from 1–3 images (Chinese tier).
8. Video Edit (video-edit): Edit an existing video with a prompt + optional reference images (Chinese tier).
9. Watermark Removal (watermark-remove): Strip Seedance 2.0 watermarks (basic or Pro).

🏷️ Tiers

Add --fast to any Global or VIP call to use the fast-queue variant (lower latency, same quality).

📥 Input Limits

Output: 4–15 seconds, auto-generated sound, 480p–720p.

⚠️ Restrictions

• No realistic human faces in uploaded images/videos (except character/omni-train modes).
• --mode extend requires a request_id from a prior seedance-v2.0-t2v or seedance-v2.0-i2v job.
• --mode first-last requires --tier global or --tier vip.
• Global/VIP omni does not support video references (images + audio only).
• --quality applies to Chinese tier only.

🔗 Core Syntax: The @ Reference System

Assign explicit roles to each uploaded asset. Tags differ by mode.

Chinese Tier (i2v, omni)

@image1  @image2  ...  @image9    (images_list order)
@video1  @video2  @video3         (video_files order)
@audio1  @audio2  @audio3         (audio_files order)

Global/VIP Omni (omni-reference-no-video / vip-omni-reference)

@image1  @image2  ...  @image9    (images_list order)
@audio1  @audio2  @audio3         (audio_files order)

Character References (all tiers)

@character:<request_id>            — from seedance-2-character or completed t2v/i2v job
@omni-character:<character_id>     — from seedance-2-omni-reference-train output

Role Assignment Table

Multi-Reference Combination

@Image1's character as the subject, reference @Video1's camera movement
and action choreography, BGM references @Audio1, scene references @Image2

🏗️ Technical Specification: The Director Brief

Structure prompts using this six-component hierarchy. Order matters — composition first, texture and micro-motion last:

Seedance 2.0 generates audio natively. Always include an Audio directive — even one sentence. Without it the model generates random ambient sound that may not match your scene.

Time-Segmented Prompts (Recommended for 10s+ videos)

Break prompts into timed segments for precise control:

0–3s: [opening scene, camera move, establishing action]
3–6s: [mid-section development, subject in motion]
6–10s: [climax or key action beat]
10–15s: [resolution, brand/product hold, text/tagline fade in]

Single-beat rule: Each segment should contain one action. 4–7s = one beat. 10–15s = 3–4 beats maximum. Overloading a segment with multiple narrative changes degrades output quality.

Negative Prompting

Seedance 2.0 supports appending negative guidance directly in the prompt. Use plain language at the end:

[your director brief above]
Avoid: camera shake, jump cuts, lens distortion, overexposure, watermarks, text overlays.

Common negative additions:

• Avoid: abrupt cuts, scene changes, multiple locations. (for single-take shots)
• Avoid: human faces, realistic people. (for product-only content)
• Avoid: fast motion, blur, unstable framing. (for smooth product reveals)

🎥 Camera Language Reference

Basic Movements

Advanced Techniques

Shot Sizes

🧠 Prompt Optimization Protocol

The Agent MUST transform user intent into a technical "Director Brief" before execution.

1. Technical Grammar: Use camera terms: Dolly In/Out, Crane Shot, Whip Pan, Tracking Shot, Anamorphic Lens, Shallow Depth of Field, High-Speed Dive, Orbital Arc.
2. Physics Directives: Use "caustic patterns," "volumetric rays," or "subsurface scattering" instead of "good lighting."
3. Timecode Notation: For multi-beat scenes, use [00:00-00:05s] format to specify timing.
4. Tag References: If files provided, use: "Replicate the camera movement of @video1 while maintaining the visual style of @image1." (lowercase, 1-based index)
5. ORDER MATTERS: Tokens at the start define composition; tokens at the end define texture and micro-motion.
6. Multi-Image i2v: Provide up to 9 reference images. The model blends aspects (style, identity, environment) across all inputs.
7. Audio is mandatory: Seedance 2.0 generates audio natively. Always include an Audio line — music genre/tone, key SFX, ambient texture. Silent direction = random audio.
8. Single-beat discipline: Each timed segment = one action. Cramming two narrative beats into 4s degrades physics and motion consistency.

🎭 Capability-Specific Patterns

1. Character Consistency

The man in @Image1 walks tiredly down the hallway, slowing his steps,
finally stopping at his front door. Close-up on his face — he takes a
deep breath, replaces the weariness with a relaxed expression.
Maintain high character consistency, zero facial flicker, persistent clothing details.

2. Camera Movement Replication

Reference @Image1's male character. He is in @Image2's elevator.
Completely reference @Video1's camera movements and facial expressions.
Hitchcock zoom during the fear moment, then orbit shots of the interior.
Elevator doors open, follow shot walking out.

3. Video Extension (Forward)

Extend @Video1 by 10 seconds.
1–5s: Light and shadow slowly slide across table through venetian blinds.
6–10s: A coffee bean drifts down. Camera pushes in toward it until screen goes black.
English text gradually appears — "Lucky Coffee", "Breakfast", "AM 7:00-10:00".

4. Video Extension (Reverse / Prepend)

Extend backward 10s. In warm afternoon light, the camera starts from
the corner with awning fluttering in the breeze, slowly tilting down
to flowers peeking out at the wall base, building anticipation for the main scene.

5. Video Editing (Modify Existing)

Subvert @Video1's plot — the character's expression shifts from warmth to
cold determination. The action is decisive, without hesitation.
Maintain all other visual elements (scene, lighting, timing).

6. Music Beat-Matching

bash scripts/generate-seedance.sh \
  --mode i2v \
  --file img1.jpg --file img2.jpg --file img3.jpg \
  --video-file reference_edit.mp4 \
  --audio-file track.mp3 \
  --subject "@Image1 @Image2 @Image3 — match the keyframe positions and rhythm of @Video1 for beat-synced cuts. BGM references @Audio1. More dynamic movement, dreamlike visual style." \
  --duration 15 --quality high

7. Dialogue / Voice Acting

In the "Cat & Dog Roast Show" — emotionally expressive comedy segment:
Cat host (licking paw, rolling eyes): "Who understands my suffering?"
Dog host (head tilted, tail wagging): "You're one to talk? You sleep 18 hours a day..."
Sound: lively studio ambience, audience laughter, punchy transitions.

8. One-Take / Long Take

@Image1 @Image2 @Image3 — one-take tracking shot following a runner
from the street up stairs, through a corridor, onto a rooftop,
finally overlooking the city. No cuts throughout.

9. E-commerce / Product Showcase

bash scripts/generate-seedance.sh \
  --mode i2v \
  --file product.jpg \
  --subject "Deconstruct the product. Static camera. Hamburger suspended mid-air, rotating slowly. Ingredients separate and reassemble. Cheese continues to melt and drip. Ultimate food aesthetics." \
  --intent "product" \
  --aspect "9:16" \
  --duration 15 --quality high

10. Science / Educational Visualization