Seedance 2.5 + MiniMax H3 提示词工具箱

简体中文（默认） · English

Seedance 2.5 与 MiniMax H3 提示词工具箱 —— 面向自媒体人、AI 视频创作者和智能体的 Agent 技能、故事板、提示词编译器与原创示例。

想直接复制使用：从 中文可复制提示词库 开始。它覆盖两模型的已核验生成模式、常见自媒体场景和平台成片结构。

这是什么

本仓库是一个中立的开源 AI 视频导演（AI video director） 层，供智能体编写 AI 视频提示词。它不是字节跳动或 MiniMax 的官方产品，不是 SDK，也不是模型托管服务 —— 仓库内没有任何模型权重、API 密钥或厂商代码。它交付的是：两个 Agent 技能（一个面向 Seedance 2.5，一个面向 MiniMax H3）、一套带 JSON Schema 的统一 VideoProjectSpec 协议、按模型区分的提示词编译器（覆盖文生视频与图生视频，以及扩展、编辑和参考模式）、一个带证据等级的来源登记表，以及 12 个经过校验的原创示例与故事板。每一条能力声明都带有来源 URL 和核对日期；第三方声明永远不会被当作官方事实呈现。

快速开始

安装两个 Agent 技能，然后可选安装编译器包。技能适用于 Codex、Claude Code、Cursor 以及任何能读取本地文件的智能体。

Codex

git clone https://github.com/xxjrq/seedance-minimax-prompts.git
cd seedance-minimax-prompts
mkdir -p ~/.codex/skills
cp -r skills/seedance-2.5 skills/minimax-h3 ~/.codex/skills/
pip install -e ".[dev]"   # 可选：编译器 + 校验器 + 测试

Claude Code

git clone https://github.com/xxjrq/seedance-minimax-prompts.git
cd seedance-minimax-prompts
mkdir -p ~/.claude/skills
cp -r skills/seedance-2.5 skills/minimax-h3 ~/.claude/skills/
pip install -e ".[dev]"   # 可选：编译器 + 校验器 + 测试

Cursor

git clone https://github.com/xxjrq/seedance-minimax-prompts.git
cd seedance-minimax-prompts
mkdir -p .cursor/skills
cp -r skills/seedance-2.5 skills/minimax-h3 .cursor/skills/
pip install -e ".[dev]"   # 可选：编译器 + 校验器 + 测试

然后直接向你的智能体提出视频需求："为 24 秒产品发布写一个 Seedance 2.5 提示词" 或 "为这个产品短片编译一个 MiniMax H3 Ref2VA 提示词"。技能会加载语法，编译器会校验 spec，最终产出可直接交给平台使用的提示词。

两个原创示例

examples/ 中每个示例都是一对文件：spec.yaml（统一的 VideoProjectSpec）和 prompt.txt（编译后的、模型专属的提示词）。下面是十二个示例中的两个，内容逐字取自仓库。

Seedance 2.5 —— 产品发布（图生视频）

输入 spec，摘自 examples/seedance-2.5/01-product-launch/spec.yaml（已精简）：

spec_version: "1.0"
project_goal: "Make a viewer want to buy the Aurora desk lamp by showing its touch-activated LED ring turning on."
model: seedance-2.5
mode: image-to-video
duration_seconds: 24
aspect_ratio: "9:16"
subject: "Aurora, a matte-black desk lamp with a warm LED ring"
scene: {setting: "a dark oak study desk in a quiet room at night", lighting: "low ambient light with a warm accent glow from the lamp ring", mood: "premium, calm, focused"}
action: "0-6s the lamp sits switched off on the desk with its ring dark; 6-12s a hand reaches in and taps the touch sensor; 12-18s the LED ring brightens from 20% to full warm white; 18-24s light spills across the desk and the lamp gently rotates toward the camera"
camera: {movement: "slow dolly-in from medium shot to close-up", angle: "slightly low angle at eye level on the lamp", lens: "shallow depth of field with soft bokeh on the bookshelf behind"}
audio: {dialogue: "no dialogue", ambient: "quiet room tone, a soft click on the sensor tap, then a faint electronic hum", music: "minimal ambient pad, swelling slightly as the lamp reaches full brightness"}
references:
  images:
    - {id: "lamp-first-frame", path: "assets/aurora-lamp-off.png", role: "first frame"}
evidence: {evidence_level: "official-doc", checked_at: "2026-08-19", source_url: "https://seed.bytedance.com/en/seedance2_5"}

编译后的提示词，逐字摘自 prompt.txt：

A 24-second cinematic video in 9:16 format, generated from the provided reference image(s).
Subject: Aurora, a matte-black desk lamp with a warm LED ring.
Scene: a dark oak study desk in a quiet room at night. lighting: low ambient light with a warm accent glow from the lamp ring. mood: premium, calm, focused.
Action: 0-6s the lamp sits switched off on the desk with its ring dark; 6-12s a hand reaches in and taps the touch sensor; 12-18s the LED ring brightens from 20% to full warm white; 18-24s light spills across the desk and the lamp gently rotates toward the camera.
Camera: slow dolly-in from medium shot to close-up. slightly low angle at eye level on the lamp. shallow depth of field with soft bokeh on the bookshelf behind. medium shot opening, close-up ending.
Performance: the hand moves slowly and deliberately, pausing before the tap. the lamp stays centered; the hand enters from the right edge.
Audio: Dialogue: no dialogue. Ambient: quiet room tone, a soft click on the sensor tap, then a faint electronic hum. Music: minimal ambient pad, swelling slightly as the lamp reaches full brightness.
Continuity: the same Aurora lamp in every frame, the same camera height and warm palette throughout.
Constraints: preserve the same Aurora lamp body; preserve the same desk; preserve the warm color palette; no text overlays; no captions; no watermarks.
Reference roles: Image 1 = first frame.

MiniMax H3 —— 多参考输入的 Ref2VA 产品短片

输入 spec，摘自 examples/minimax-h3/05-ref2va-product-multi-ref/spec.yaml（已精简）：

spec_version: "1.0"
project_goal: "Show a matte-black desk lamp turning on, combining product stills, a motion reference, and a hum track into one short product clip."
model: "minimax-h3"
mode: "ref2va"
duration_seconds: 6
aspect_ratio: "1:1"
subject: "a matte-black desk lamp with a warm LED ring"
references:
  images:
    - {id: "lamp-front", path: "assets/lamp-front.png", role: "subject"}
    - {id: "lamp-detail", path: "assets/lamp-detail.png", role: "style"}
  videos:
    - {id: "lamp-motion", path: "assets/lamp-motion.mp4", role: "motion reference"}
  audio:
    - {id: "lamp-hum", path: "assets/lamp-hum.mp3", role: "soundtrack"}
constraints:
  preserve: ["the matte-black finish", "the warm LED ring"]
  forbidden: ["no brand logos", "no text overlays"]
evidence: {evidence_level: "official-doc", checked_at: "2026-08-19", source_url: "https://github.com/MiniMax-AI/MiniMax-H3"}

编译后的提示词，逐字摘自 prompt.txt：

subject_definitions:
<Subject 1> is the visual content referenced from the provided image (asset lamp-front, role: subject), guiding the appearance and style of the target video.
<Subject 2> is the visual content referenced from the provided image (asset lamp-detail, role: style), guiding the appearance and style of the target video.
<Video 1> is the reference video (asset lamp-motion, role: motion-reference), providing motion, camera, and temporal structure for the target video.
<Audio 1> is the reference audio (asset lamp-hum, role: soundtrack), providing the audible layer for the target video.

summary:
[reference generation + audio reuse] The target video shows a matte-black desk lamp with a warm LED ring in a home office at night with a tidy oak desk, lit by low ambient light with a warm accent glow, with a calm mood.

retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved - the referenced appearance and style are retained in the target video.
<Subject 2> (appears in [Shot 1]): fully_preserved - the referenced appearance and style are retained in the target video.
<Video 1> (motion and structure): weak_reference - the target video follows the reference video's motion and pacing without copying its frames.
<Audio 1>: reference - the target video references the audio's character without copying the original signal.

detailed_description:
The target video is in a realistic, cinematic style with natural lighting and a coherent color palette.
[Shot 1] a home office at night with a tidy oak desk, lit by low ambient light with a warm accent glow, with a calm mood. a matte-black desk lamp with a warm LED ring. a hand taps the touch sensor, the lamp brightens from dim to full, and the light ring casts a soft halo on the desk. The camera arcs slowly around the lamp at eye level with a 50mm lens in a close-up. <Subject 1> and <Subject 2> appear as the referenced visual content.
Continuity must preserve: the matte-black finish; the warm LED ring.
Forbidden: no brand logos; no text overlays.

overall_soundscape: a soft click, then a faint electrical hum.

non_diegetic_music: a minimal ambient pad.

能力对比

表格中每一格都标注了官方来源，核对日期为 2026-08-19。这里没有任何第三方声明。

H3 其他官方规格：24 FPS 输出、11 种稳定对话语言（阿拉伯语、中文、英语、法语、德语、意大利语、日语、韩语、葡萄牙语、俄语、西班牙语）、BF16 CFG-distilled 检查点、首发版仅支持全注意力推理。Seedance 2.5 官方页面还列出了视频编辑、白模控制、绿幕编辑、专业运镜和表演调度（performance blocking）。分辨率、帧率和输出格式在 Seedance 2.5 官方页面未公布，因此本仓库不作声明。

仓库结构

seedance-minimax-prompts/
├── README.md                  # 中文主入口
├── README.en.md               # 英文版 README
├── llms.txt                   # 面向 LLM 的仓库索引
├── SKILL.md                   # 顶层路由技能（seedance-minimax-prompts）
├── CONTRIBUTING.md            # 如何添加示例与更新事实
├── SECURITY.md                # 安全策略
├── CHANGELOG.md               # 版本历史
├── LICENSE                    # MIT —— 代码、Schema、脚本
├── LICENSE-CONTENT            # CC BY 4.0 —— 文档、示例、故事板
├── NOTICE                     # 第三方署名与不转载政策
├── THIRD_PARTY_SOURCES.md     # 已审核的第三方仓库及其声明
├── schemas/                   # JSON Schema：video-project、shot-plan、reference-manifest、generation-request
├── models/                    # 按模型区分：capability.yaml、constraints.yaml、prompt-grammar.md
├── sources/                   # official-sources.yaml + third-party-providers.yaml（带 verified_at 日期）
├── src/ai_video_director/     # router.py、validators.py、source_registry.py、compilers/
├── skills/                    # seedance-2.5/ 与 minimax-h3/ Agent 技能
├── examples/                  # 12 个黄金示例（每模型 6 个）：spec.yaml + prompt.txt
├── evals/                     # 评测用例与预期输出
├── tests/                     # pytest 测试套件（80 个测试）
└── docs/                      # architecture.md、license-audit.md

提示词编译器

一套统一协议，每个模型一个编译器。编译契约如下：

VideoProjectSpec (schemas/video-project.schema.json)
  → validate_spec()            校验 Schema + 各模型硬约束
  → route()                    按 spec["model"] 选择编译器
  → model-specific compile()   Seedance 2.5 或 MiniMax H3 提示词语法
  → validate_generation_request()  再次校验 Schema + 各模型约束
  → GenerationRequest          可直接交给平台的输出

• schemas/video-project.schema.json 是唯一的模型无关输入：project_goal、model、mode、duration_seconds、aspect_ratio、subject、scene、action、camera、audio、references、constraints、evidence。
• 编译器位于 src/ai_video_director/compilers/ —— seedance_25.py（文生视频、图生视频、视频扩展、视频编辑）与 minimax_h3.py（文生视频、首帧、尾帧、首尾帧、Ref2VA）。
• 本仓库不提供 CLI，直接调用 route()：

from ai_video_director.router import route

request = route(spec)          # spec: 符合 video-project.schema.json 的 dict
print(request["prompt"])       # 模型专属、符合语法的提示词

违反约束时抛出 ai_video_director.validators.ValidationError，并附可操作的错误信息 —— 任何数值都不会被静默截断或取整。模型能力绝不跨模型继承：Seedance 的模式或限制不会被假定适用于 MiniMax H3，反之亦然。

校验与证据理念

仓库中每条动态事实都带证据块：evidence_level（official-doc / official-demo / third-party / unverified）、checked_at（ISO 日期）和 source_url。能力文件还记录 surface（official-web / official-api / local-model），让读者清楚事实来源。

• 官方数据与第三方数据绝不混用。第三方提供商的参数只存在于 provider 块和 sources/third-party-providers.yaml 中，永远不会出现在官方能力文件里。
• 仅出现在第三方包装器中的声明 —— Seedance 原生 4K、72 条 API 路由、放宽审核、定价、端点、"完整本地开源 H3 2K 管线"、"官方 SDK" —— 绝不当作官方事实呈现。完整登记表见 THIRD_PARTY_SOURCES.md，许可结论见 docs/license-audit.md。
• 来源会按计划重新核对；source_registry.check_recency() 会标记 verified_at 缺失或过期的条目。

参与贡献

见 CONTRIBUTING.md —— 如何添加黄金示例（spec 必须通过校验，prompt 必须往返一致）、如何更新能力事实（必须更新 checked_at 并保留 source_url），以及 PR 检查清单。

许可证

• 代码、Schema 与脚本：MIT —— 见 LICENSE。
• 文档、提示词示例与故事板：CC BY 4.0 —— 见 LICENSE-CONTENT。
• 第三方署名与不转载政策：见 NOTICE。

本项目与字节跳动、MiniMax 无关联，不受其背书，也不是其产品。MiniMax H3 相关材料受 MiniMax Community License 约束；本仓库不复制 MiniMax 的源代码、权重或任何受许可限制的内容。

相关链接

• 字节跳动 Seedance 2.5 官方产品页面 —— https://seed.bytedance.com/en/seedance2_5
• MiniMax H3 官方仓库 —— https://github.com/MiniMax-AI/MiniMax-H3
• MiniMax H3 Community License Agreement —— https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
• MiniMax 开放平台 API 文档（视频生成）—— https://platform.minimax.io/docs/api-reference/video-generation-v2-create