跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
视频与音频

seedance

Expert guide for prompting ByteDance's Seedance 2.0 AI video model (text-to-video, image-to-video, first/last-frame, multi-reference, and multi-shot generation with native synchronized audio). Use when the user wants to write, improve, or debug a Seedance 2.0 prompt; plan a shot; work with reference images/videos/audio; animate a start frame or interpolate first-and-last frames; direct camera moves, motion, and cinematic style; generate dialogue, sound effects, or music; keep a character consistent across shots; or understand Seedance 2.0's capabilities, specs, API model IDs, limits, and access routes (Dreamina, Volcano Ark, BytePlus, fal.ai, Replicate). Triggers on "Seedance", "Seedance 2.0", "ByteDance video model", "Dreamina video", "seedance prompt", "reference-to-video", "first and last frame video".

maystudios · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add maystudios/claude-skills --skill seedance

技能说明


name: seedance description: > Expert guide for prompting ByteDance's Seedance 2.0 AI video model (text-to-video, image-to-video, first/last-frame, multi-reference, and multi-shot generation with native synchronized audio). Use when the user wants to write, improve, or debug a Seedance 2.0 prompt; plan a shot; work with reference images/videos/audio; animate a start frame or interpolate first-and-last frames; direct camera moves, motion, and cinematic style; generate dialogue, sound effects, or music; keep a character consistent across shots; or understand Seedance 2.0's capabilities, specs, API model IDs, limits, and access routes (Dreamina, Volcano Ark, BytePlus, fal.ai, Replicate). Triggers on "Seedance", "Seedance 2.0", "ByteDance video model", "Dreamina video", "seedance prompt", "reference-to-video", "first and last frame video".

Seedance 2.0 — Prompting & Capabilities

Seedance 2.0 (ByteDance Seed team, launched Feb 2026) is a multimodal audio-video model: it takes text, images, audio, and video in a single call and generates a 4–15s clip with native synchronized audio (dialogue + lip-sync, sound effects, ambient, music) in the same pass.

Core mental model — this drives every decision below: Seedance 2.0 is a physics and cinematography simulator you direct like a film crew, not a mood board you brainstorm with adjectives. It renders physical interactions ("tires smoke as the car pivots on wet asphalt") and directorial language ("slow dolly-in", "golden-hour rim light"). It does not render vague adjectives ("cinematic", "epic", "amazing"). Write shot directions, not vibes.

Capabilities at a glance

ModeInputUse it for
Text-to-video (T2V)prompt onlyGenerating a scene from scratch
Image-to-video (I2V)1 start image + promptAnimating an existing still
First-and-last frame2 images (first_frame+last_frame) + promptControlled transitions / reveals / before-after
Reference-to-video (R2V)1–9 images + ≤3 videos + ≤3 audio (max 12 files) + promptCharacter/style/motion/object consistency, voice, "omni" composition
Multi-shotprompt with a shot list / timelineA short narrative (up to ~6 shots) in one 15s generation
Video-extendprior clip + promptContinuing past 15s by chaining

Native specs: 480p/720p native (1080p and native 4K on hosting platforms), 4–15s, 24 fps, aspect ratios auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. Audio is on by default (generate_audio: true), at no extra cost. Full specs, access routes, model IDs, pricing, and the raw API request shape are in references/access-and-specs.md — read it when the user asks how to call/access the model or needs exact limits.

The universal prompt formula

Order matters: the model reads left-to-right and weights the front most heavily. The first 2–3 instructions land reliably; adherence drops after ~8 requirements. Front-load subject and action.

[Subject] + [Action] + [Environment] + [Camera] + [Lighting] + [Style] + [Constraints]
  • Subject — who/what, with concrete material/age/wardrobe detail
  • Action — ONE clear action, present tense (compound choreography causes "melting")
  • Environment — where, time of day, weather
  • Camera — ONE primary move + shot size (see rules below)
  • Lighting — the single highest-leverage element; always name it explicitly
  • Style — 2–3 named anchors (director/film-stock/movement), never vague adjectives
  • Constraints — what to keep fixed, plus duration & aspect ratio

Target 60–100 words for a single shot (hard ceiling ~3000 characters; the model starts averaging past ~150 words). Write natural, directorial prose — not a keyword list.

Worked example:

"A weathered fisherman in a yellow raincoat hauls a net over the side of a small wooden boat. Grey choppy sea, dawn, light drizzle. Camera slow tracking shot from the water's surface, gentle rise. Hard rim light through overcast sky, desaturated teal grade, 35mm film grain. Realistic water weight and rope tension, feet planted. 8 seconds, 16:9. Avoid jitter and bent limbs."

Non-negotiable rules (the highest-leverage ones)

  1. One primary camera move per shot. "Slow push-in" ✅. Stacking pan + zoom + dolly + handheld ❌ → "soup".
  2. Never mix camera motion and subject motion in one clause. ✅ "The dancer spins slowly. Camera holds a fixed frame." ❌ "spinning camera around a dancing person" → jitter.
  3. Name lighting explicitly. If you add only one thing to improve a prompt, add a lighting description.
  4. Physics, not adjectives. Describe weight, friction, impact, cloth/fluid behavior. Never write bare "fast" (the #1 quality-killer) — qualify pace ("slow, smooth, gradual") and apply speed to one element only.
  5. Named style anchors beat vague words. "Wes Anderson symmetry, 35mm Kodak grain" ✅. "beautiful, cinematic, epic" ❌.
  6. State consistency locks in any reference-based or multi-shot prompt, or characters/outfits drift.
  7. Scale beats to duration (multi-shot): ~2–3 beats/5s, 4 beats/10s, 5–6 beats/15s. One action + one camera move per beat.

These plus the full camera/motion/style/multi-shot/negative-prompt vocabulary and 10 worked example prompts are in references/prompting.md — read it for any non-trivial prompt authoring, camera direction, multi-shot storyboard, or when the output has artifacts to debug.

Reference tagging basics

When any image/video/audio is attached, bind it into the prose with @Image1, @Video1, @Audio1 (numbered by input order) and always state what each reference controls — identity vs. camera-motion vs. style vs. audio. Vague "reference @Video1" is a top failure mode.

  • ✅ "The woman in @Image1 with dark curly hair walks down a neon street; reference @Video1 for the camera movement only; background music from @Audio1."
  • ❌ "Woman, @Image1, dancing, @
阅读完整来源 ↗