Seedance 2.0/2.5 Prompt Engineering

Seedance is ByteDance's quad-modal video model: image + video + audio + text input. Output: up to 1080p, 4–15s per generation (Seedance 2.0), extendable to longer cuts via stitched segments (Seedance 2.5), auto dubbing/scoring, native lip-sync. Prompts are high-density storyboards, not one-line descriptions.

When to Use This Skill

• User asks to write / improve a Seedance 2.0 prompt
• User wants to port a shot idea into Seedance structure
• User mentions multi-shot storyboards, camera moves (Hitchcock zoom, FPV, bullet time), character consistency across shots, video-to-video edits, or audio-driven generation
• User requests a specific style: cinematic, anime, commercial, UGC, short-drama (竖屏), VFX, POV/bodycam, retro found-footage (home video/VHS/MiniDV), pets/animals as lead, stop-motion, game-UI/livestream HUD, fashion lookbook, food ASMR, horror/suspense, process/transformation montage, or storyboard-grid-to-video

Core Prompt Anatomy

Every strong Seedance 2.0 prompt combines 5 blocks. Order is flexible but all blocks matter.

[Style]      → genre, film ref, director ref, render engine, resolution, color grade
[Duration]   → total seconds (4–15s); Seedance caps at 15s
[Scene]      → setting, time of day, weather, atmosphere
[Shot List]  → timecoded blocks ([00-05s], [05-10s], [10-15s]) with shot type,
                camera move, subject action, micro-details, dialogue cue
[Tech/Neg]   → aspect ratio, fps, lens, lighting; negatives (optional 排除: block)

Minimum viable template

Style: {genre/director ref}, {resolution}, {color grade}.
Duration: {N}s.

[00-Ts] Shot 1 ({shot type}): {camera move}. {subject} {action}.
  Details: {micro-detail}. Dialogue Cue: "{line}"
[T-2Ts] Shot 2 ...
[2T-Ns] Shot 3 ...

Tech: {aspect ratio}, {fps}, {lens}, {lighting}.

Multi-Modal Reference Syntax

Seedance accepts explicit reference tags inline. Use sparingly — one ref per asset.

Reference chaining in a single prompt is supported — e.g. identity from image, camera from video, music from audio: 参考@图1的男人形象，他在@图2的电梯中，参考@视频1的运镜.

Three Generation Modes

Category Catalog

Detailed templates and examples for each category live in references/.

• references/patterns.md — structural patterns + idioms mined from 1000+ prompts and verified cases
• references/templates.md — fill-in templates per category (cinematic, anime, commercial, UGC, short-drama, VFX, sound, consistency, camera, extension, POV/bodycam, found-footage, pets, stop-motion, game-UI, fashion, food-ASMR, horror, process montage, storyboard-grid-to-video)
• references/camera-vocab.md — camera moves, lens, lighting, grading, POV optics, retro-defect, stop-motion and HUD vocabulary
• references/examples.md — curated working prompts with source attribution

Quick Rules

1. Always timecode shots. [00-05s] blocks → Seedance's temporal alignment is shot-based, not descriptive.
2. Three shots is the sweet spot for 15s. Five max. More → drift.
3. One action per second. Don't pack; over-dense shots desynchronize.
4. Director/film refs work. Denis Villeneuve style, Wong Kar-wai 90s Hong Kong, Hans Zimmer score → strong tonal anchors.
5. Dialogue cues inline with lip-sync intent: Dialogue Cue: "line" (200-400ms pauses).
6. Aspect ratio matters. 9:16 for vertical short-drama / TikTok, 2.35:1 for cinematic, 16:9 default.
7. Negative prompts use 排除: (Chinese-native pattern) — 排除：模糊，低清，噪点，水印，文字，扭曲变形，五官崩坏.
8. Character consistency → repeat identity anchor in each shot (同一人物, maintain consistent face/clothing/hair).
9. For extension/continuity → end shot N with exact state, start shot N+1 from that state (pose, lighting, framing).
10. Audio drives rhythm — if using @audio, beat-align actions (beat-synced, 与音乐节拍完全吻合).
11. Lock identity once, not per shot. State the identity/wardrobe lock in one fixed block and add a whole-clip "maintained throughout" clause; repeating it inside every shot header actually triggers drift instead of preventing it.
12. Closed, summing timecodes. Segments must touch end-to-end (0-4s, 4-8s...) and sum exactly to the stated duration — a stated duration with gaps makes the model stretch the last beat.
13. One dialogue line ≤ ~8 words per 2-5s beat, one speaker per beat. Longer lines or two simultaneous speakers break lip-sync.
14. End each beat with an explicit end-state line (pose/expression/position) so the next beat has a defined starting point — critical for multi-shot continuity and for dialogue reaction chains.

Anti-Patterns

• One-line descriptions → Seedance under-performs vs shot-list format
• Mixing Chinese and English mid-sentence without commas → parser trips
• More than 2 reference tags per shot → identity conflict
• Abstract emotion-only prompts (no visual action) → static output
• Over-specifying > 8 micro-details per shot → model drops some
• Restating wardrobe/hairstyle inside every shot → appearance mutation between shots (lock once, reference the lock)
• Writing a duration without segmenting it → visible drift after ~6 seconds on unsegmented clips
• Asking for "realistic handheld" + "cinematic 4K lighting" in the same prompt → two contradictory light logics, result looks plastic
• Expecting the model to render clean on-screen logos/slogans/longer type → leave a clean tail frame, add text in post

Workflow

1. Ask user: style, duration, assets on hand (images/videos/audio)
2. Pick closest template from references/templates.md
3. Fill 3-shot storyboard with timecodes
4. Add style header + tech footer + (optional) negative block
5. If references exist, insert @图片N / @视频N tags
6. Output prompt in a single fenced block — user pastes into Seedance