SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsseedance-prompt

seedance-prompt

Design & creation

View repository

Updated 10-16 05:48

About this skill

Seedance 2.0 Prompt Writing Skill

This Skill is distilled from the official Jimeng Seedance 2.0 user manual to help users write high-quality Seedance 2.0 video generation prompts.

1. Platform Parameter Quick Reference

Dimension Limit
Image input ≤ 9 images
Video input ≤ 3, total duration ≤ 15s
Audio input MP3, ≤ 3, total duration ≤ 15s
Total mixed input limit 12 files
Generation duration 4-15s, freely selectable
Audio output Built-in sound effects/music (native audio-visual sync)

2. Two Entry Modes

2.1 First and Last Frame Mode

  • Applicable to: uploading only the first frame image (or first frame + last frame) + text prompt
  • Scenarios: simple image-to-video, first-last frame transition animation

2.2 All-Purpose Reference Mode

  • Applicable to: scenarios requiring multimodal combined input (image + video + audio + text)
  • Interaction method: specify the purpose of each asset via @素材名 Example:
  • Example: @Image1 as the first frame, @Video1 for reference camera language, @Audio1 for background music@图片1 作为首帧，@视频1 参考镜头语言，@音频1 用于配乐

3. Core Prompt Formula

Seedance 2.0 prompts follow one core formula:

素材角色指定 + 动作/剧情描述 + 镜头语言 + 氛围/音效指令

3.1 Asset Role Assignment (WHO / WHAT)

Use the @ symbol to clearly define the role of each asset in the frame. Key principles:Each asset must have a clear identity and purpose。

Common assignment patterns:

Pattern Writing example
Assign as character @图片1的男人 / 参考@图片1@图片2长枪角色
Assign as first frame @图片1作为首帧画面 / @图片1作为画面的首帧图
Assign as last frame @图片2作为尾帧
Assign as scene 他在@图2的电梯中 / 在@图5的枫叶林中
Assign as prop/product 对@图片2的包包进行商业化摄像展示 / 手中提着@图片3的包
Assign as style reference 包包的表面材质参考@图片3
Assign as motion reference 模仿@视频1的动作 / 完全参考@视频1的所有运镜效果
Assign as audio reference 旁白的音色参考@视频1 / 背景BGM参考@视频3中的音效
Assign as rhythm reference 视频节奏参考@视频 / 参考@视频中的画面关键帧的位置和整体节奏
Assign as appearance but not first frame @图片1的女子走到镜子前 / 参考@图1的男人形象

Multi-asset combination example:

参考@图1的男人形象，他在@图2的走廊中，完全参考@视频1的所有运镜效果
服装参考@图片1@图片2的样式，手中提着@图片3的包，视频节奏参考@视频

3.2 Action/Plot Description (WHAT HAPPENS)

Seedance 2.0's biggest advantage isDescribe in narrative language, and the model automatically plans the storyboard。

Efficient description principles:

  1. Narrate along the timeline rather than using technical terms

    • ✅ 男人下班后疲惫的走在走廊，脚步变缓，最后停在家门口
    • ❌ 生成一个行走动画，逐渐减速

  2. Actions should be specific and visualizable

    • ✅ 快速的将手伸出画框拿起可乐喝了一口，然后露出一脸满足的表情
    • ❌ 角色拿起饮料

  3. Emotions should be explicitly written out

    • ✅ 脸部特写镜头，男人深呼吸，调整情绪，收起了负面情绪，变得轻松
    • ✅ 沉思了一会突然开始崩溃大叫
    • ❌ 角色情绪变化

  4. Dialogue should be written directly in the prompt (supports multilingual lip sync)

    • ✅ 用普通话说"做完他的，做你的，大家文明排队。"
    • ✅ 女主说话"何人擅闯我永宁侯府？"

  5. Use timestamps to segment complex plots

    0-3秒画面：[场景A描述]
3-6秒画面：[场景B描述]
7-12秒画面：[场景C描述]
13-15秒画面：[结尾画面描述]

3.3 Camera Language (HOW TO SHOOT)

Seedance 2.0 supports two camera control methods:

Method A: Reference video camera movement (recommended, most precise)

完全参考@视频1的所有运镜效果
参考@视频1的运镜、画面切换节奏
运镜参考@视频1，推近到屏幕的特写，镜头旋转后反转展示全貌

Method B: Text-described camera movement (used when there is no reference video)

Shot type Prompt writing
One continuous shot 一镜到底的追踪镜头，从街头跟随...上楼梯、穿过走廊、进入屋顶
Push in/macro 镜头前推 / 推近到特写 / 镜头推进特写茶杯
Pull out 镜头拉远展示酒馆全景 / 镜头小幅度拉远（露出街头全景）
Follow 镜头跟随黑衣男子快速逃亡 / 正面跟拍 / 侧面跟拍
Orbit 几个环绕镜头展示电梯内视角 / 近的环绕镜头
Pan shot 镜头右摇 / 镜头快速向右边摇动 / 急停后右摇180度
High-angle shot 转为空中的俯拍视角 / 半空中俯拍
Subjective perspective 第一人称视角 / 主观视角 / 从第三人称视角变成人物的主观视角
Static shot 固定镜头
Hitchcock zoom 主角在惊恐时希区柯克变焦
Cut 切镜，特写 / 镜头切到空中的俯拍视角

Shot size description:

  • 脸部特写镜头 / 手部特写 / 中景 / 全景 / 远景

3.4 Atmosphere/Sound Effects Instructions (SOUND & MOOD)

Seedance 2.0's native audio capabilities are very strong, and can be controlled directly in the prompt:

Sound effect control:

背景音效为走路声，人群声，汽车声
背景音恢宏大气
分镜切换以及重点的情节演绎加入特殊音效
出现电吉他的激昂音乐

Narration/dialogue control:

用科普风格和音色，将图片1中的内容演绎出来
全程自然对话
音色参考@视频1

Style and atmosphere:

整体风格诙谐幽默
电影级写实风格，采用2.35:1宽银幕，24fps，细腻的画面风格
黑白水墨风格

4. Seven Major Scene Templates

4.1 Product Advertisement

对@图片N的[产品]进行商业化的摄像展示，[产品]的[角度A]参考@图片N，
[产品]的[材质/细节]参考@图片N，要求将[产品]的细节均有所展示，
背景音[氛围描述]

Advanced version (with scene narrative):

[时间戳分段]
0-Ns画面：[场景建立，产品入场]
N-Ns画面：[产品特写/功能展示]
N-Ns画面：[使用场景/情感连接]
最后N秒画面：[品牌收束，产品全貌+slogan]

4.2 Character Narrative Short Film

@图片1的[角色]在[场景]，[一系列连贯动作描述]，
全程自然对话/旁白，[情绪氛围]

4.3 Action Reference Replication

参考@图1@图2[角色A]，@图3@图4[角色B]，
模仿@视频1的动作，在@图5的[场景]中[行为]

4.4 Creative Special Effects Replication

将@视频1的[人物/首帧]替换成@图片1，
完全参考@视频1的特效和动作，[特效变化描述]

4.5 One-Shot

@图片1@图片2...@图片N，一镜到底的[镜头类型]镜头，
从[起点]...[中间节点]...最终[终点]

4.6 Video Extension

将@视频1延长Ns。[新增内容的时间戳分段描述]
注意：生成时长选择 = 新增部分时长

4.7 Video Editing/Subversion

颠覆@视频1里的剧情，[新剧情描述]
将@视频1中的[角色A]换成[角色B/图片]，[其他保持不变的要素]

5. Prompt Optimization Checklist

After writing the prompt, check it against the following list:

  • Clear asset annotations: Each @asset has a clear purpose (character/scene/action reference/first frame, etc.)
  • Specific and visual actions: Avoid abstract descriptions; use language that can be "filmed"
  • Reasonable timeline: The amount of content within 15 seconds is moderate (usually 3-5 key actions/scene transitions)
  • Emotions are annotated: The character's expressions, emotions, and tone are clearly described
  • Camera has direction: At least specify whether it is follow/push in/static/one-shot
  • Sound effects are implied: Describe ambient sound, dialogue content, or specify the music atmosphere
  • Reasonable number of assets: Mixed input does not exceed 12 files
  • Duration matches content: The amount of content matches the selected generation duration (information density should not be too high)

6. Common Pitfalls and Best Practices

Writings to avoid

  • ❌ Piling up pure technical parameters (such as "4K, HDR, 60fps" — the model may not respond to all parameters)
  • ❌ Cramming descriptions with more than 15 seconds of content into one prompt
  • ❌ Vague @asset annotations (such as only writing "refer to these images" without saying the role of each image)
  • ❌ Dialogue content is too long (audio synchronization has length limits; each dialogue segment should be short and powerful)

Recommended practices

  • ✅ First use a 5-second short video to test the prompt effect and product fidelity, then generate 15 seconds after satisfaction
  • ✅ Use Jimeng to first generate reference images (such as character design sheets, scene images), then input them as assets
  • ✅ When there is a reference video, prioritize using "reference @video's camera movement" rather than textual description of camera movement
  • ✅ In product advertisements, add "all details are shown" to improve product fidelity
  • ✅ For scenes with dialogue, specify the language (Chinese/English/dialect); the model supports multilingual lip synchronization
  • ✅ For complex scenes, use timestamp segmentation (0-3s / 3-6s / ...) to let the model understand the rhythm more precisely
  • ✅ You can directly write the text/subtitle content appearing in the frame in the prompt