tutorial · 1 min read

一句提示词 → 一支竖屏 AI 舞蹈短片

用 GPT Image 2 生成一张带标注的舞步网格,交给 Seedance 2 Pro 作为参考,让编舞真正跟着动作走,再拼接成一支卡点音乐的竖屏短片。

大多数 AI 舞蹈视频看起来很糟,是因为模型在逐帧猜测动作。解决办法是先用 GPT Image 2 预先编排一张动作网格,把它作为参考图传给 Seedance 2 Pro——这样它就有了可循的分镜脚本。这个配方把网格、舞蹈和匹配的音乐床串联成一支短片。

#cli#claude-code#pipelines#dance#seedance#gpt-image-2#reel#automation
Run it

Two ways, same result

Either paste the prompt into your AI agent, or run a single command in your terminal: both invoke the same recipe.

est. cost ≈ 14.2 credits needs PIPE2_TOKEN runs 6 pipelines
prompt · paste into agent
Claude Code, Codex, any shell-agent: conventions in AGENT.md
Run the pipe2 dance-reel recipe via the pipe2 CLI. Defaults produce a K-pop breakdancer reel with a vocal-led Eleven Music bed: pass --subject, --style, --moves, --music to customize, or --persona / --watermark-url for branded output. Report the final video URL when done.
The agent reports the final hosted video URL when the chain completes.
command · run in shell
Needs pipe2 CLI + PIPE2_TOKEN; fetches the recipe from GitHub
pipe2 recipe run dance-reel
Prints the hosted video URL to stdout when the chain finishes.

为什么有效

网格就是分镜脚本。image-generator 以 4×4 布局渲染 16 个带标注的 舞蹈姿势,Seedance 把它读作参考图——按顺序跟随每个格子,而不是 逐帧猜测动作。输入一个命名动作,输出的就是同一个动作。

音乐是在舞蹈之前创作的,而非之后。Seedance 会根据成品音轨对 片段进行卡点,因此编舞会踩在节奏上,而不是与后期叠加的音乐床各走各的。

关键的旋钮

  • moves —— 提升质量最省力的杠杆。命名动作(“太空步、 身体波浪、驴踢”)胜过泛泛的形容词:每个名称都会变成一个带标注的 网格格子,而 Seedance 会读取这些标注。
  • dance_style —— 为网格和舞蹈同时定调(“K-pop 独舞编舞”、 “古典芭蕾”)。要具体;“跳舞”不是一种风格。
  • subject / persona —— subject 是谁在跳舞、穿着什么; persona 锚定一个反复出现的身份,让同一张脸和同一套服装贯穿整个 片段。具体的着装和场景比泛泛的人物形象更出彩。

时长、音乐情绪、水印和画幅比例都有合理的默认值——完整选项参见清单。

备注

自带输入是成对出现的:文本标志用于生成素材,-url 变体用于提供 你自己的素材。persona / persona-url、音乐情绪 / music-url、 水印提示词 / watermark-url——用提示词生成,用 URL 自带。拥有固定 logo 或精选音轨的品牌就用 -url 那一侧。

Walkthrough

6 pipelines transform one input into one output

What each step takes in, and what it spits out. The artifact glyph on the right shows the output kind.

total ≈ 14.2 credits steps 6 pipelines flow

Estimated at recipe defaults. Final cost may vary: metered models bill by actual token usage, and optional steps add to the total when their inputs are supplied.

  1. GPT Image 2 renders a 16-panel dance move reference grid: one labeled pose per cell.

    image
  2. Eleven Music composes a vocal-led K-pop dance bed. Generated FIRST so the dance can be choreographed to its rhythm.

    audio
  3. image-motion 0.5 cr

    Claude designs a ken-burns pan over the grid so it reads on screen before the dance starts.

    video
  4. Seedance 2 Pro generates the dance clip. The grid tags as @image1 for choreography, the music tags as @audio1 for beat-conditioning: choreography is synced to the actual track.

    video
  5. video-reel 1 cr

    Claude stitches grid-reveal + dance with a snappy crossfade and the music bed.

    video
  6. watermark 0.5 cr

    Overlays the Pipe2 logo (or your own --watermark-url) at the top-left corner so the clip carries attribution across reposts. Pass --no-watermark to skip.

    video

Click any slug to see full pipeline pricing tiers.

See also

Stuck on a chain that won't compose? Drop into our Discord and paste the JSON output of the failing step.