视频模型 Prompt 指南:Seedance vs 可灵 vs Sora

更新于 2026-11 · 阅读约 6 分钟

同一个镜头描述,投给不同模型出来的画面差异巨大——因为各模型对提示词的理解偏好不同。这篇按平台拆解写法,全部可以在 Prompt 构建器里一键生成。

中文平台

Seedance 2.5(即梦 / 小云雀)

字节系旗舰,支持多模态参考(最多 50 种输入类型)和 30 秒连续生成。写法要点:主体 + 动作 + 镜头 + 光线 + 风格五要素齐全;角色一致性优先用"上传角色参考图"而不是纯文字描述;需要跨镜头一致时明确写"与参考图完全一致"。

可灵 Kling 3.0

物理运动真实感是强项,动作、打斗、光影特效。写法要点:突出动作的物理逻辑("衣摆随转身摆动""雨水沿窗滑落"),角色一致性表现优秀,适合动作向与奇幻向短剧。

海螺 Hailuo(MiniMax)

速度快、单条成本低,适合批量试错。写法要点:描述精炼聚焦一个核心动作与情绪,堆太多元素反而容易崩。

海外平台

OpenAI Sora

动作连贯性与叙事理解最强。英文提示词建议结构:Cinematic quality + [shot type], [camera move]: [content]. Lighting... Style... ~Xs. 先声明画质基调再描述内容。

Runway Gen-4

风格化控制最灵活,镜头语言 token 丰富。适合艺术向、风格化项目,明确写 camera language。

Google Veo 3.1

提示词遵循度最佳,支持原生音频——可以连台词口型和环境音一起生成,短剧场景非常省一步配音。写法加一句 Audio: ...。

角色一致性:三个层次

  1. 文字锁定:固定的可视化标签(年龄 + 发型发色 + 服装 + 标志性道具),每个镜头 prompt 末尾原样复用
  2. 参考图锁定:Seedance / 可灵支持上传角色设定图,比文字强得多
  3. 工作流锁定:在结构化 JSON 的 characters 里存 visualTags,所有下游提示词自动引用(见剧本工程化)

分镜分级:省 40% 渲染成本

不是所有镜头都值得高配抽卡。把镜头分三档:

一部 80 集的剧,C 级镜头通常占一半以上,这一项就能省下约 40% 渲染成本。

负面提示词模板

中文平台:避免:画面闪烁、人物面部变形、多余的手指、字幕水印、镜头快速抖动
海外平台:Avoid: flickering, face morphing, extra fingers, burned-in subtitles or watermarks, shaky camera

想直接生成?打开 Prompt 构建器,选平台、填描述、勾上角色锁定,复制即用。

Video Model Prompt Guide: Seedance vs Kling vs Sora

Updated Nov 2026 · ~6 min read

The same shot description produces wildly different results across models — each has its own prompt preferences. Here's the per-platform breakdown; every pattern below can be generated in one click in our Prompt Builder.

Chinese platforms

Seedance 2.5 (Jimeng / Xiaoyunque)

ByteDance flagship with multimodal reference (up to 50 input types) and 30-second continuous generation. Structure: subject + action + shot + lighting + style. For consistency, prefer uploading a character reference sheet over text-only description, and state "exactly consistent with the reference image" explicitly.

Kling 3.0

Strongest physics-aware motion: action, combat, light and shadow effects. Emphasize physical logic ("clothes swing with the turn", "rain slides down the window"). Excellent character consistency — a good pick for action and fantasy dramas.

Hailuo (MiniMax)

Fast and cheap per clip — ideal for volume trial-and-error. Keep descriptions tight: one core action, one emotion. Over-stuffed prompts break it.

Western platforms

OpenAI Sora

Best narrative understanding and motion continuity. Recommended English structure: Cinematic quality + [shot type], [camera move]: [content]. Lighting... Style... ~Xs. Declare the quality baseline before content.

Runway Gen-4

The most flexible stylized control with rich camera-language tokens. Best for artistic and stylized projects — be explicit about camera language.

Google Veo 3.1

Best prompt adherence and native audio — dialogue with lip-sync plus ambience in one generation, which saves an entire dubbing step for short drama. Append an Audio: ... line.

Character consistency: three levels

  1. Text lock: a fixed visual tag (age + hair + outfit + signature prop) reused verbatim at the end of every prompt
  2. Reference-image lock: Seedance / Kling accept character sheets — far stronger than text
  3. Workflow lock: store visualTags in your structured JSON's characters and let every downstream prompt reference them (see Script Engineering)

Shot tiering: save ~40% render cost

Not every shot deserves max settings. Tier your shots:

In an 80-episode show, Tier C usually exceeds half of all shots — this alone cuts render cost by roughly 40%.

Negative prompt template

Avoid: flickering, face morphing, extra fingers, burned-in subtitles or watermarks, shaky camera

Want it generated for you? Open the Prompt Builder, pick a platform, describe the shot, check the consistency lock, and copy.