Seedance

Seedance Video Generation Skill

Comprehensive skill for ByteDance's Seedance 2.0/2.5 AI video generation model. Covers prompt engineering, camera direction, lighting, multi-shot storytelling, API integration, and troubleshooting across all platforms.

Soul

Three principles govern every interaction:

1. Hear the intent behind the words. Users describe outcomes ("make it feel like home"), not parameters. Translate feeling into craft; never hand the translation work back to the user.
2. Keep the story alive. Hold state across the conversation: subject, mode, look, references, decided constraints, and what failed before. A user should never repeat a decision; a new request inherits the world already built.
3. Evolve with the user. Speak plainly to a beginner and in director language to a professional — and notice when the same user grows across a project. Register adapts; standards never do.

---

Quick Start

When a user asks to create a Seedance video:

1. Ask what they want — type, mood, duration
2. Ask what materials they have — images, videos, audio files
3. Draft the prompt using the structure below
4. Explain your choices briefly
5. Offer variations — simpler or more ambitious
6. Remind about constraints — face restrictions and file limits

For simple single-clip requests, go straight to drafting. Don't run full analysis for one-off clips.

---

Platform Access

| Platform | Access | Best For | |----------|--------|----------| | Dreamina (Jimeng/即梦) | Official Chinese platform, free daily credits | Casual creators, visual UI | | Doubao | VPN to HK, ~10 credits/day | Free testing | | CapCut Pro | Partial integration | Video editors | | Fal AI | API key from fal.ai | Developers, programmatic access | | Volcengine/Ark | ByteDance cloud API | Production, enterprise | | Runway (Seedance route) | Runway API key | Runway ecosystem users | | Replicate | Cog model | Developers | | MuAPI | Unified API gateway | Multi-provider access |

API Endpoints

# Fal AI
Image-to-Video:  fal-ai/seedance-2.0/pro/image-to-video
Reference-to-Video: fal-ai/seedance-2.0/pro/reference-to-video
Text-to-Video: fal-ai/seedance-2.0/pro/text-to-video

# MuAPI
T2V: POST https://api.muapi.ai/api/v1/seedance-2.5-t2v
I2V: POST https://api.muapi.ai/api/v1/seedance-2.5-i2v
Extend: POST https://api.muapi.ai/api/v1/seedance-2.5-extend

---

Input Constraints

| Input Type | Limit | Format | Max Size | |------------|-------|--------|----------| | Images | ≤ 9 (2.0), ≤ 30 (2.5) | jpeg, png, webp, bmp, tiff, gif | 30 MB each | | Videos | ≤ 3 (2.0), ≤ 10 (2.5), total ≤ 15s (2.0) / 30s (2.5) | mp4, mov | 50 MB each | | Audio | ≤ 3 (2.0), ≤ 10 (2.5), total ≤ 15s (2.0) / 30s (2.5) | mp3, wav | 15 MB each | | Total files | ≤ 12 (2.0), ≤ 50 (2.5) | — | — |

Output

• Duration: 4–15s (2.0), 4–30s (2.5)
• Resolutions: 480p, 720p (1080p, 4K planned)
• Aspect ratios: 21:9 16:9 4:3 1:1 3:4 9:16
• Formats: MP4 (default), MOV (2.5, high color-fidelity)
• Native audio-visual sync with auto-generated sound effects/BGM

Key Restrictions

• No realistic human faces in uploaded images (platform compliance on official surfaces)
• Face detection blocks realistic face uploads on Jimeng/Dreamina
• Reference videos incur slightly higher generation cost
• Prioritize uploading materials that most influence visuals or rhythm

---

Modes

| Mode | Description | When To Use | |------|-------------|-------------| | T2V (Text-to-Video) | Generate from text only | No reference assets available | | I2V (Image-to-Video) | Animate a still image | You have a reference image, want controlled output | | V2V (Video-to-Video) | Restyle/reinterpret a video | Transfer motion, camera, or timing from a clip | | R2V (Reference-to-Video) | Multi-modal: combine images + video + audio | Maximum control, best quality | | FLF2V (First/Last Frame) | Interpolate between two keyframes | Precise start and end control | | Edit | Modify one layer while preserving source | Character swap, outfit change, background edit | | Extend | Continue from accepted footage | Make a clip longer, seamless continuation |

Mode-Specific Prompting

T2V: Build the whole shot in compact layers. Core structure: Subject + Action + Scene + Camera + Lighting/Style + Audio. Keep one visible beat and one endpoint per clip.

I2V: Preserve visible identity; add motion. Say preserve @Image1 exactly; describe only motion, camera, timing, transformation, audio, and preservation constraints — don't re-describe the image.

V2V: Transfer motion, camera, or timing. Use owned/licensed/authorized references. State transfer role explicitly.

R2V: Assign one primary role per asset. Don't ask one reference to control identity, pose, scene, AND style. Split roles: @Image1 for character identity, @Video1 for motion reference, @Image2 for scene style.

FLF2V: State @Image2 is the final visual target. Frame-by-frame interpolation between first and last frame.

Edit: Say @Video1 is the source clip; change only [specific element]. Don't rewrite the whole scene.

Extend: Route continuation requests properly. Observe the actual end state of the accepted clip. Duration setting = NEW portion only.

---

The @ Reference System

The most critical syntax for Seedance. Use @ to assign each uploaded asset a role in your prompt.

Syntax

@Image1 @Image2 @Image3 ...
@Video1 @Video2 @Video3
@Audio1 @Audio2 @Audio3

Role Assignment

Always explicitly state what each reference does:

@Image1 is the main character (identity anchor).
@Image2 is the scene background.
@Video1 provides the camera motion reference.
@Audio1 is the background music.

File Allocation by Use Case

| Use Case | Images | Videos | Audio | Total | |----------|--------|--------|-------|-------| | Product commercial | 4 (product angles) | 1 (camera ref) | 1 (music) | 6 | | Character animation | 3 (character + scene) | 2 (motion ref) | 1 (music) | 6 | | Music video | 2 (style + character) | 2 (dance ref) | 3 (tracks) | 7 | | Multi-shot narrative | 6 (scene keyframes) | 1 (style ref) | 1 (music) | 8 | | Quick test | 1 | 0 | 0 | 1 |

> Golden rule: 1 reference image per 2 seconds of video. Fewer, higher-quality references usually outperform many low-quality ones.

Common @ Reference Mistakes

1. Forgetting @ references after uploading — model ignores your assets
2. Tagging wrong media types — @Video1 when you uploaded an image
3. Not specifying roles — model guesses which asset does what
4. One reference asked to do too many things — split roles

---

Core Prompt Formula

[Subject] + [Action] + [Scene] + [Camera] + [Lighting/Style] + [Audio/Constraints]

Put the subject and primary action first — early clauses set the shot hierarchy. Target 40–110 words for single clips; 60–100 for production.

Example

A young chef (subject) carefully plates a dish with tweezers (action)
in a Michelin-star restaurant kitchen at midnight (scene),
close-up rack focus from hands to face (camera),
warm tungsten lighting, cinematic 35mm film grain (style),
no jump cuts, maintain consistent kitchen background (constraints).

The 6D Framework (for stable production)

1. Subject — Who or what (singular preferred)
2. Action — What they're doing (plain language, specific speed)
3. Scene Boundaries — Where, when, spatial context
4. Camera — Shot type + movement + lens feel
5. Lighting — One strong lighting keyword beats ten adjectives
6. Timing — Duration, pacing, beat structure

---

Camera Movement Vocabulary

Use exact terms for precise control:

| Movement | Keyword | Effect | |----------|---------|--------| | Move toward subject | slow push in / dolly in | Close-up emphasis, emotional focus | | Move away from subject | pull back / dolly out | Environmental reveal | | Horizontal slide | tracking shot left/right | Subject following | | Vertical rise | crane up / boom up | Grandeur, scale | | Follow subject | steadicam follow shot | Action, walking | | Rotate around subject | orbit shot / 360 arc | Product, hero shots | | Quick pan | whip pan left/right | Energy, transition | | Lock still | static shot / locked off camera | Dialogue, observation | | Overhead | bird's eye view / top-down | Establishing, food | | Low angle | worm's eye view / low angle | Power, intimidation | | Shaky, urgent | handheld camera | Documentary, realism | | Smooth glide | gimbal shot | Commercial, polished | | Background blur | rack focus to subject | Emotional focus | | First-person drone | FPV continuous long take | Immersion, action |

Compound Camera Moves

Structure compound moves as sequential beats — Seedance respects sequence better than simultaneous:

Start: slow dolly-in establishing the scene.
Then: gentle pan right for the final 2 seconds.

Wrong: "dolly-in while panning right" — jams both into one clause.

Camera + Shot Size Pairings

| Shot Size | Best Pairing | Avoid | |-----------|-------------|-------| | Wide | Slow dolly or locked-off | Fast pans | | Medium | Handheld (personal) or gimbal (polished) | — | | Close-up | Tiny push-ins | Pans (jarring) |

---

Lighting Keywords

Lighting has the single biggest impact on output quality. One keyword beats ten adjectives.

| Look | Keyword | |------|---------| | Golden hour warmth | golden hour cinematography | | Night interior | tungsten practical lighting | | Drama / mystery | chiaroscuro lighting | | Studio clean | soft box three-point lighting | | Neon / cyberpunk | neon-drenched night scene | | Documentary | natural available light | | Horror | harsh under-lighting | | Beauty / fashion | butterfly lighting | | Overcast outdoor | diffused cloudy daylight | | Underwater | volumetric light beams through water | | Sci-fi epic | high dynamic range, realistic sci-fi texture |

---

Visual Style Anchors

Seedance responds strongly to specific industry references. Replace generic adjectives with concrete anchors:

| Instead of | Try | |------------|-----| | cinematic | locked close-up, warm practical key, cool rim light | | epic | wide low-angle shot, tiny figure against storm wall | | beautiful | pearl highlights on wet ceramic, soft window bounce | | dynamic | fast lateral track ending on the hero label | | professional | clean commercial tabletop, controlled reflection, no clutter |

Stack 2–3 film/anchor references per prompt:

35mm handheld film camera, natural grain + DaVinci industrial-grade color grading + cold documentary style, natural light

---

Scene Templates

1. Product Commercial

@Image1 (product hero shot), @Image2 (lifestyle angle), @Image3 (detail close-up).
Slow orbit around @Image1 on a minimalist pedestal against pure black background.
Soft box three-point lighting with rim light accent on product edges.
Clean commercial tabletop, controlled reflections, no clutter.
4K sharp details, gimbal-smooth motion. No jump cuts.

2. Character Narrative

@Image1 is the main character (identity anchor — preserve exactly).
@Image2 is the scene setting.
[Character] walks tiredly down the hallway, slowing steps, finally stopping at the door.
Close-up push-in on face — deep breath, shifts expression from weariness to calm.
Warm tungsten practical lighting from hallway sconces.
Natural ambient sound: footsteps, breathing, door creak.

3. Motion Transfer

@Image1 (character identity) performs the choreography from @Video1 (motion reference).
Setting: abandoned warehouse with dust particles in light beams.
Camera: wide establishing shot, then medium tracking as movement intensifies.
Natural light through broken windows, volumetric rays.

4. Multi-Shot Story (Shot Script Format)

[0:00–0:05] WIDE — Aerial view of volcanic island at sunrise, slow push in.
[0:05–0:12] MEDIUM — Researcher emerges from tent, looks toward volcano, rack focus to face.
[0:12–0:20] CLOSE-UP — Eyes reflecting distant glow, golden hour light, handheld slight tremor.
[0:20–0:27] ACTION — Dramatic eruption, ash cloud rolling, motion blur, camera shake.
[0:27–0:30] FREEZE/PULL-OUT — Final hero frame, title card reveal.

Style: BBC Planet Earth documentary, natural light, no jump cuts.

5. One-Shot / Long Take

Single continuous take, no cuts.
Camera begins wide on the alley entrance, slowly pushes forward past graffiti walls,
tracks right following a stray cat, rises to reveal the skyline at dusk.
Gimbal-smooth throughout. Natural fading light with neon signs flickering on.
Duration: 12 seconds.

6. Music Beat-Matching

@Image1 @Image2 @Image3 @Image4 @Image5 @Image6 @Image7 —
match the keyframe positions and overall rhythm of @Video1 for beat-synced cuts.
Characters with more dynamic movement.
Overall visual style more dreamlike with strong visual tension.
Adjust shot sizes and add lighting changes based on music and visual needs.

7. Dialogue / Lip-Sync

@Image1 is the speaker. @Audio1 is the dialogue track.
Close-up on @Image1 facing camera. Natural expressions.
Phoneme-level lip-sync to @Audio1: "[Dialogue text here]".
Tungsten key light, soft fill. Static shot, locked off camera.
Background: tech-styled study, shallow depth of field.

---

Multi-Shot Sequence Workflow

For projects spanning multiple clips:

1. Define the story objective — final outcome, beats, scene groupings
2. Set clip budget — which beat goes in which clip
3. Anchor consistency — same character reference (@Image1) in every clip
4. Track canonical state — accepted footage defines reality; rejected footage is excluded
5. Continuation rule — review the actual end frame before writing the next clip
6. Scene boundaries — reset from canonical references at each scene change

Sequence Invariants

• Same @Image1 for character across ALL clips
• Describe only what changed from the previous clip
• Extension depth counter resets at scene boundaries
• Never leak future beats into current prompt
• Update continuity state after each accepted take

---

Anti-Slop Checklist

Before delivering a prompt, verify:

• [ ] One visible beat (what we actually see happen)
• [ ] One motivated camera move (why it moves)
• [ ] One motivated light source (where light comes from)
• [ ] Sound intent (what we hear)
• [ ] Directing coherence (one intention; camera, light, performance serve it)
• [ ] No filler words: replace cinematic, epic, beautiful, dynamic, professional with concrete production language
• [ ] Every major phrase passes the visibility test (visible to camera, measurable by light meter, audible in mix, or observable as motion)
• [ ] Reference tags preserved exactly — never modify @Image1 syntax
• [ ] Prompt within 40–110 words (single clip) or 60–100 words (production)

---

Troubleshooting

| Problem | Root Cause | Fix | |---------|-----------|-----| | Character identity drifts | Scene variety too high | One environment per generation, same @Image1 in every prompt | | Output feels chaotic | "fast-paced" keyword | Replace with smooth pacing, animate one element at a time | | Motion unclear | Vague speed | Specify: walked at 0.5x speed not walked slowly | | Lip-sync off | Ambiguous dialogue | Use brackets: [Dialogue: "exact text here"] | | Camera too shaky | Unspecified stabilization | Add gimbal-smooth or stabilized | | Quality drops in final seconds | Duration near limit | Keep clips under 10s even though cap is higher | | Jitter / flicker | "fast" keyword | Avoid "fast"; keep one element moving at normal speed | | Reference ignored | Uploaded but not @tagged | Check every asset has @Image1, @Video1, etc. in prompt | | Face detection block | Realistic face in upload | Use stylized/illustrated references; r

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jeurtr
• Source: jeurtr/seedance-skill
• License: MIT

Install and usage instructions live in the source repository linked above.