MiniMax H3 提示词指南说的不是一句氛围感描述，而是一份紧凑的拍摄脚本：告诉模型画面里发生什么、镜头如何运动、谁在对白、以及 4–15 秒片段里立体声如何配合画面。
在 KreadoAI 上，最高效的路径是 H3 拍摄脚本栈：先选对输入模式（纯文本、首帧、首尾帧或全参考），再按 MiniMax 官方字段块撰写，最后用两行声场字段锁定音效。这套结构与 H3 原生立体声输出、2K 导出，以及单条最高 7,000 字符的提示词上限完全对应。
本指南将 MiniMax 官方提示词语法，适配电商钩子、品牌英雄镜头与多模态外观锁定等广告团队场景，全部落在 KreadoAI MiniMax H3 工作流内。
为什么 H3 提示词不同于「电影感」一句话
MiniMax H3 面向多模态上下文理解而设计，不是抽奖式氛围词。官方方法把每次生成都当作导演简报：
| 氛围词写法 | H3 期望写法 |
|---|---|
| 形容词堆叠 | 按 [Shot N] 标记的分镜时间线 |
| 「电影感环绕」 | 具体运镜：Push In、Truck Left、Static Shot |
| 音效事后补写 | 两个必填字段：overall_soundscape + non_diegetic_music |
| 上传参考但不定义角色 | 全参考模式：六个带保留规则的结构化段落 |
对付费社交和 DTC 团队而言，实际变化是：你可以在 8–12 秒钩子里同时锁定产品外形、镜头节奏和原生立体声氛围——但前提是提示词读起来像脚本，而不是 Pinterest 文案。
若需先了解产品能力，请访问 MiniMax H3 模型页。本文是提示词写作指南手册。
H3 拍摄脚本栈 — 基础模式与全参考模式
MiniMax 根据是否上传参考素材，将提示词分为两套栈结构。
基础栈 — 纯文本、首帧或首尾帧（3 个字段）
在 KreadoAI 上运行文生视频、图生视频（首帧）或首尾帧，且未使用全参考上传时，使用此栈。
可选首行 — 帧对齐说明（有静帧锚点时才粘贴）：
text For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
首尾帧模式使用：
text How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 8.00-second mark of the target video.
随后是三个核心字段：
text integrated_multimodal_description: [Shot 1] Live-action, cinematic, ...
overall_soundscape: ...
non_diegetic_music: ...
字段职责一览：
| 字段 | 控制内容 |
|---|---|
integrated_multimodal_description |
画面、动作、分镜、对白、画内音 — 主时间线 |
overall_soundscape |
全片环境音与物理动作声（不含对白） |
non_diegetic_music |
观众可听、角色不可听的背景音乐 — 无音乐时写 N/A |
全参考栈 — 多模态参考模式（6 个字段）
在 KreadoAI 全参考工作流中上传最多 9 张图、3 段视频、3 段音频（合计 12 个文件）时，MiniMax 要求六个段落：
text subject_definitions: is the amber serum bottle in
summary: [reference generation + audio reference] The target video shows on white seamless with a macro drop hook, 9:16 paid social, 10 seconds.
retention_analysis: (appears in [Shot 1]): fully_preserved - bottle shape, label, and cap color locked.
detailed_description: The target video uses a clean beauty commercial style with soft studio key light. [Shot 1] A medium close-up establishes on white seamless. The camera pushes in with small amplitude at slow speed as a golden drop falls in macro slow motion...
overall_soundscape: Quiet studio room tone with a soft liquid drip.
non_diegetic_music: Sparse piano at a slow tempo, fading at the end.
全参考是 H3 对广告团队最有价值的模式：从参考素材继承角色外观、镜头节奏与音色，同时用结构化字段让漂移可见、可修。
动笔前先选对 KreadoAI 模式
H3 提示词结构取决于输入类型。模式错配会导致比例错误或关键帧被忽略。
| KreadoAI 工作流 | MiniMax 任务 | 对齐首行 | 适用场景 |
|---|---|---|---|
| 文生视频 | T2VA | 无 — 直接从三核心字段开始 | 概念测试、UI 动效、分镜预演 |
| 图生视频（首帧） | I2VA | <Picture 1> 在 0.00s |
电商主图转动态 |
| 首尾帧 | FL2VA | Picture 1 在 0.00s，Picture 2 在 S.SS | 产品从开盖到英雄定格 |
| 全参考 | Ref2VA | 六字段全参考栈 | 品牌外观锁定、角色+产品+音频参考 |
| 视频生视频 | V2V 编辑/续写 | Omni 栈中的 <Video N> |
动作迁移、片段修复 |
KreadoAI 侧需提前规划的约束：
- 提示词必填，最高 7,000 字符
- 时长：4–15 秒（对齐行中的结束时间写为两位小数
S.SS） - 分辨率：768P 草稿或 2K 成片
- 首尾帧图生视频：比例跟随上传静帧，不可单独选比例
- 文生视频与全参考：支持自适应、16:9、9:16 等
运镜、对白与声场 — 写作规则
以下规则来自 MiniMax 官方指南，对广告提示词影响最大。
运镜 — 类型 + 幅度 + 速度
在分镜内用自然英文描述运镜，不要堆叠标签：
text The camera pushes in with small amplitude at slow speed toward the product label. The camera trucks right with large amplitude at fast speed, revealing the full pack shot. The camera holds a static shot as steam rises from the mug.
避免只说「电影感环绕」，除非明确写出 Arc Shot 及幅度、速度。
分镜与切换
[Shot 1] 不加时间戳。后续分镜在选定时长内使用严格递增的切换时间：
text [Shot 2] At 00:04.500, the camera cuts to a close-up of the product cap...
硬切优先用 the camera cuts to。只有明确需要时才写 cross-dissolve 或 fade。
对白与画外音
为说话者分配稳定 ID — (S1)、(S2) — 并用标签包裹台词：
text The presenter with a warm mid-range voice (S1) says: <d>[English] Fresh glow, every day.</d>
画外音使用 says in an off-screen voiceover，并说明画内角色嘴唇保持闭合。<d> 内保留用户原始语言，不要自行翻译。
声场字段 — 不可省略
即使计划导出静音版付费社交素材，也要写满两行：
text overall_soundscape: Quiet studio room tone with faint product handling sounds. non_diegetic_music: N/A
对白与画内音乐写在 integrated_multimodal_description 或 detailed_description 中 — 不要写进 overall_soundscape。
在 KreadoAI 运行 H3 提示词
- 打开 KreadoAI MiniMax H3，或在 图生视频 中选择 MiniMax H3。
- 选定模式 — 纯文本、首帧、首尾帧或全参考 — 按最终顺序上传参考素材。
- 粘贴 H3 拍摄脚本栈：对齐行（如有）→ 核心或全参考字段 → 声场两行置后。
- 时长对齐脚本（钩子 8–10 秒、中层 12–15 秒），分辨率草稿用 768P、成片用 2K，比例对齐投放位。
- 生成后以 1× 速度审听原生立体声，每次只改一个字段迭代。
迭代原则：每次重跑只改一个段落。产品变形 → 修 subject_definitions 或首帧锚点。节奏不对 → 修分镜切换时间。音频漂移 → 修 overall_soundscape / non_diegetic_music。
需要 30 秒级时间戳脚本？参见 Seedance 2.5 提示词指南 — H3 更擅长 4–15 秒带立体声的多模态广告片段。
可直接投放的提示词示例
H3 提示词跨品类复用同一套拍摄脚本栈 — 变的只是主体、场景与声场。以下 7 组可复制模板，覆盖美妆、服装、食品、3C、家居、首尾帧与全参考工作流。
| 行业 | 示例 | 模式 | 比例 | 适用场景 |
|---|---|---|---|---|
| 美妆 / 护肤 | A | T2VA | 9:16 | 精华钩子、质地特写 |
| 服装 / 鞋履 | B | T2VA | 9:16 | 球鞋上新、面料细节 |
| 食品 / 饮料 | C | T2VA | 9:16 | 气泡感、原料英雄镜头 |
| 消费电子 | D | T2VA | 16:9 | 产品揭幕、发布会能量 |
| 家居 / 家电 | E | T2VA | 16:9 | 厨房场景、功能演示 |
| 家居 / 香氛 | F | FL2VA | 跟随静帧 | 从包装到氛围定格 |
| 宠物 / 生活方式 | G | 全参考 | 9:16 | 角色 + SKU 锁定 |
示例 A — 美妆 / 护肤 9:16 钩子（10 秒，T2VA）
text integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium close-up frames an amber glass serum dropper bottle on a white seamless backdrop. The camera pushes in with small amplitude at slow speed as a single golden drop falls in macro slow motion, catching soft studio key light. Liquid shimmer rolls across the glass surface while the bottle shape and label remain unchanged. The shot holds with the product centered and the top fifteen percent of frame clear for headline overlay.
overall_soundscape: Quiet studio room tone with a soft liquid drip and faint glass resonance.
non_diegetic_music: Sparse ambient piano at a slow tempo with gentle fade at the end.
示例 B — 服装 / 鞋履 9:16 上新（10 秒，T2VA）
text integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium shot frames white-and-orange running sneakers on a cream-to-coral gradient studio backdrop. The camera trucks left with small amplitude at slow speed as the laces ripple slightly and sole tread catches rim light. The lens pushes into a brief texture macro on mesh upper and stitching before returning to a hero hold with both shoes centered and the top fifteen percent of frame clear for headline overlay.
overall_soundscape: Soft studio room tone with faint fabric rustle and a light sole tap on the surface.
non_diegetic_music: Upbeat lo-fi groove at a moderate tempo, steady volume, no swell.
示例 C — 食品 / 饮料 9:16 英雄镜头（8 秒，T2VA）
`text integrated_multimodal_description: [Shot 1] Live-action, cinematic, a close-up frames a chilled glass bottle of sparkling citrus drink with condensation on a warm terracotta surface. The camera pushes in with small amplitude at slow speed as bubbles rise through amber liquid and the cap catches window light. A fresh orange slice rolls into the lower edge of frame. The shot holds with the bottle centered, label area clean, and the top fifteen percent clear for headline overlay.
overall_soundscape: Faint glass contact sound, soft fizz crackle, and light surface scrape as the orange slice settles.
non_diegetic_music: Bright acoustic guitar at a moderate tempo with gentle fade at the end.
示例 D — 消费电子 16:9 英雄镜头（10 秒，T2VA）
text integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a sleek matte-black wireless earbuds charging case on a dark gradient pedestal in a minimal studio. The camera trucks right with small amplitude at slow speed, revealing the case from profile to three-quarter angle. Soft rim light traces the metal edge while product geometry stays locked. The camera holds a static shot for the final two seconds with the product centered and the bottom twenty percent clear for offer overlay.
overall_soundscape: Low room tone with a faint surface contact sound as the case settles.
non_diegetic_music: Restrained electronic pulse at a slow tempo, fading at the end.
示例 E — 家居 / 家电 16:9 生活方式（10 秒，T2VA）
`text integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a matte-white electric kettle on a light oak kitchen counter with morning window light and a soft plant blur in the background. The camera pans right with small amplitude at slow speed as steam begins to rise from the spout and a warm amber indicator light glows. The camera holds a static shot for the final two seconds with the kettle centered and the bottom twenty percent clear for offer overlay.
overall_soundscape: Warm domestic room tone, gentle kettle hum, soft steam hiss.
non_diegetic_music: N/A
