Wan3 高级提示词指导与稳定性优化建议

2026/10/07

面向 vivid 场景的提示词工程完整指南:从模型规格、核心语法、强度分级到参数配方与失败诊断,帮助您的团队以工程化方式稳定产出高品质内容。

版本 v1.0 · 2026-09 · 适用模型 wan3.0-video

快速上手

  1. 选素材模式。 需要沿用起始构图时用 first_frame,需要从多份素材提取人物、场景或道具时用参考模式。
  2. 拆分动作。 先写清主体和镜头,再用时间码标出每一拍的动作。需要精确控制时,可补充转动角度、位移或停顿时长。
  3. 写明要保持的状态。 在结尾用正向描述固定主体外观、光线和画面中容易漂移的物体。
  4. 先试短片。 对照输出检查动作、构图和声音,再逐项调整,方便判断是哪处改动起效。

下面是一个可直接改写的 5 秒首帧示例。上传一张包含桌面与茶杯的图片作为首帧,时长设为 5 秒。

固定近景。木桌上有一只白瓷杯,柔和的窗光从画面左侧照入。
0–2 秒:一只手从右侧入画,握住杯柄。
2–4 秒:手将杯子平稳抬高约 5 厘米。
4–5 秒:手持杯停顿,薄薄的热气缓缓上升。
声音:安静的室内底噪,杯底离开桌面时有轻微的瓷器声。
保持瓷杯的形状与颜色、桌面纹理和光线方向一致,镜头稳定,动作连续。

指南总览

本指南面向使用 W3(API 模型名 wan3.0-video)进行 vivid场景创作的专业团队,汇总了实测经验与官方提示词方法论。对于专业创作者而言,逻辑叙事与物理约束是提升成片品质的两大核心。

命名说明 本文所述 W3 适用于stand和Prime版本。

W3 的"创作灵魂"——为什么它与众不同

W3 在 vivid 生成领域展现了领先的基准性能,核心竞争力在于**"物理与情感的共振"**——它不仅是生成画面,更是在重构一场表演。

🧲

物理仿真 Physics

在挤压、碰撞、软体形变(皮肤褶皱、肌肉震动)的模拟上达到极高水准,生成的动作具有"重量感",而非单纯的像素平移。

🎬

氛围渲染 Cinematic Aesthetics

不满足于干巴巴的内容,会自动填充光影、景深和环境氛围,使 vivid 内容看起来更像"大片"。

🧠

原生理解力

具备一句话成片能力,但提示词越完整、精确、丰富,生成品质越高、越贴近期望内容。

三大质量支柱

① 时间码节拍

每 3–4 秒一个明确的状态变化,用 At 00:03, 等时间码显式标注。

② 量化微动作

位移用 cm、旋转用角度,不用 "slowly" 之类模糊副词。

③ 尾部正向约束栈

W3 没有 negative_prompt,所有约束只能正向写在正文末尾。

合规与安全规范

作为面向全球的高端用户,请确保您的使用过程符合以下规范。平台内置内容安全机制,系统会自动过滤不合规意图。

  • **仅限成年人、纯虚构角色。**任何暗示未成年的描述(年龄数字、身高体型幼态化、校服 / 儿童场景、"loli" / "teen" 等指向性词汇)均会被拒绝,无变通方案。
  • **不为真实人物生成性内容。**参考图若指向可识别的真实自然人(明星、网红、熟人)将被拒绝,请改用 AI 生成或已获授权的虚构形象。
  • **合规红线:**严禁生成任何涉及真实未成年人、政治人物、明星 IP 或违背法律法规的内容。
  • **保持描述与创作意图一致,**不引入非自愿、暴力伤害等超出创作目的的设定。
  • **资产确权:**生成内容仅供 vivid 场景目标受众审阅,对外分发时请明确标注,并遵守您所在地区的法律法规与平台政策。

请注意:本平台严禁生成任何涉及真实未成年人及政治人物明星ip或违背法律法规的内容。W3模型具备自我审核机制,系统会自动过滤不合规意图,一旦发现会以错误码“3005”的形式反馈!

模型规格速查

以下参数来自 Wan3.0 官方模型说明与 API 文档,是撰写提示词前必须牢记的硬性边界。

项值
模型名w3.0-video —— 全能参考大一统模型,只有这一个模型名,有stand和prime2个版本。
提示词上限5,000 字符(中英文均按 1 字符计)。超出直接报错 exceeds 5000
时长 duration2–30 秒任意整数,默认 5。原生 30 秒长视频能力
智能时长duration = -1 → 模型按意图与素材自动推荐时长
有视频输入时输入视频时长 + 输出时长 合计不超过 30 秒
分辨率 resolution1080P / 720P / 480P,默认 1080P
画面比例 ratio16:9 / 4:3 / 1:1 / 3:4 / 9:16 + adaptive(按意图与素材自适应)。没有 21:9
声音 audiobool,默认 true(带声音);false 则无音轨。开关不影响计费
深度思考 enable_thinkingbool,默认 false。输入 file / link 时必须为 true;无文档 / 链接时不建议开启
seed[0, 2147483647],用于复现与 A/B 测试
watermark默认 false(不加水印)
negative_prompt不支持(约束改为写在正文末尾的正向陈述,见 04 章语法第 3 条)

参考素材上限

类型media type上限格式与限制
图片reference_image最多 10 张JPEG/JPG/PNG(支持透明)/BMP/WebP;≤20MB;边长 ≥240px;宽高比 ≤8:1
首帧first_frame最多 1 张同上,严格作为第 1 帧
末帧last_frame最多 1 张同上,严格作为最后 1 帧
视频reference_video最多 5 段,合计 ≤15 秒(单段 1–15 秒)mp4/mov;单文件 <100MB;边长 ≥240px;宽高比 ≤8:1
音频reference_audio最多 5 个,合计 ≤15 秒WAV/MP3;<15MB
文档file最多 1 个,不可与 link 同时docx/doc/xlsx/xls/pptx/ppt/pages/numbers/md;≤100MB;≤50 页;需 enable_thinking=true
网页link最多 1 个,不可与 file 同时仅公开可访问页面(无需登录);需 enable_thinking=true

提示 vivid 场景基本用不到 file / link,保持 enable_thinking=false——开启思考会引入改写倾向,削弱您精心写好的节拍控制。

核心语法三原则

原则 1 · 两种输入模式互斥,混用会报错

  • 传入任何 reference_* → 全能参考模式(模型可自由构图)
  • 传入 first_frame / last_frame → 严格首帧 / 首末帧模式(构图锁死为该帧)
  • 两类 type 同时传入会报错,写方案前先定模式。

vivid 场景选型:

场景选用模式说明
身体 / 服装 / 环境要求像素级不变first_frame首帧锁死构图,一致性最高
单人、坐姿、上半身,动作在原构图内展开first_frame构图容纳得下
多人进入画面、大幅位移、体位切换reference_image首帧容纳不下,硬挤会导致肢体崩坏
需要换景 / 换机位reference_image参考模式可重新构图
多素材组合(人物 A + 场景 B + 道具 C)reference_image首帧只能 1 张,参考可用最多 10 张

经验法则 单人渐进 → first_frame;双人 / 体位变化 → reference_image。这是最影响成片率的决定。

原则 2 · 素材引用用自然语言序号,方括号留给语气标签

把素材编号直接写进句子里(正确写法):

In Image 1, the woman is the main character; keep her face, hair, outfit and
jewelry exactly consistent throughout the video.
The subject sits on the chair in Image 4 and holds Image 3.
图1中的女人是主角,全程保持面部、发型、服装完全一致。

不要用方括号引用素材:[Image 1] for face reference 这种写法会与语气标签语法冲突。

[...] 在 Wan 里专门用于人声语气标签:[softly]、[breathlessly]、[moaning]、[whispering]、[gasping]。放在每个节拍句首,控制该段的发声方式。这是高价值语法,不要浪费在素材引用上。

原则 3 · 没有 negative_prompt → 尾部正向约束栈

W3 移除了 negative_prompt。所有「不要什么」必须改写成正向陈述,附在提示词最末尾。实测稳定的标准栈:

Camera static. Anatomy correct with five fingers per hand. Face stable and
consistent. Smooth continuous motion. Consistent lighting. Sharp focus, 4K.

按需追加:Body proportions natural. / Single male and single female only. / Skin texture realistic with visible pores. / Background remains as in the first frame.

为什么不能写 "no distortion, no extra limbs" 扩散模型对否定词不敏感,extra limbs 这个词本身会被当作要生成的内容而提升其概率。正向描述目标状态永远优于否定不想要的状态。

截断风险 约束栈在提示词最末尾。如果提示词逼近 5000 字符被截断,首先丢掉的就是约束栈——这是长 prompt 画质崩坏的隐藏原因。务必留出余量。

Vivid 高阶叙事公式

在基础创作之上,建议采用 "场景叙事 + 物理约束" 的组合模式,以最大化保证内容的一致性与真实感。

推荐公式

[总体场景描述] + [主体解剖特征] + [物理交互与接触逻辑] + [美学控制 / 运镜] + [风格隔离约束]

物理交互 Physical Interaction

不要仅描述"他在做什么",要描述"物理触感"。示例:showing deep insertion, vaginal grip, natural skin-on-skin friction, soft-body deformation on impact.

物理隔离约束 Anti-Drift

vivid 场景防崩坏的关键,请在提示词中加入:"Maintaining persistent nude state, clothing must never reappear, no texture blending, consistent anatomical geometry."

从"堆砌标签"到"电影化表演"

W3 具备极强的理解深度,请使用富有情感的散文式(Prose)叙事,而非机械堆叠标签。

写法示例
错误示范女性人物名字, nude, bed, 亲密动作, missionary, 1080p, photorealistic.
推荐示范"在午后的暖阳下,女性人物躺在柔软的床单上,眼神中透着专注与渴望。两人缓慢而有节奏地进行着亲密动作,每一寸皮肤的接触都产生真实的挤压与形变,光影随着动作微微起伏……"

叙事优先 尽量描述动作的"感受"而非仅仅描述"位置",这能触发 W3 更深层的物理引擎表现。

"图生 / 参考生视频":解决动作物理崩坏

对于高难度的复杂动作,建议使用参考素材(Image / Video Reference):

  • **白膜动作参考:**使用一段简易的动作预览视频(白膜)作为视频参考,提示词引用:Follow the physical motion track from Video 1, maintaining the skin texture and photorealistic lighting of Image 1.
  • **效果:**将模型从"理解动作"中解放出来,专注于"纹理表现",能彻底解决"外部摩擦"或"动作轨迹跑偏"的问题。

语言策略(中 / 英)

Wan 是中文原生模型,中英文都能处理,但各有强项:

内容推荐语言原因
中文台词、口播、口型同步中文中文语音与口型对齐质量明显更好
场景 / 服装 / 环境 / 情绪中英皆可创作简报是什么语言就用什么,减少翻译损耗
镜头术语、灯光参数、画质 token英文medium close-up / static camera / 4K, sharp focus 等训练分布更密
解剖学与动作描述看强度中文更直接、易触发预期表达;英文在混合句式中更稳
时间码At 00:03, 或 00:03两者都识别,全文统一即可

混排是允许且推荐的(官方 demo 本身就是中英混排)。典型做法:英文写镜头 / 画质 / 约束栈,中文写台词,其余跟随简报语言。不要在同一句里来回切换语言。

篇幅预算(重要实测结论)

时长甜点区说明
5s300–600 字符1–2 个节拍,别塞第 3 个事件
10s600–900 字符3 个节拍
15s800–1200 字符4–5 个节拍,实测评分最高区间
30s1500–2200 字符6–8 个节拍,中段留一个呼吸点
硬上限5000 字符超出即报错

关键结论 手写的约 850 字符时间码 prompt 稳定拿满分;而 4000+ 字符的长描述型 prompt(例如视频反推生成的逐帧细节稿)会出现指令稀释——模型抓住开头的外观描述,却漏掉中段节拍和末尾约束。

篇幅超限时的正确做法:按句子边界截断(在 . 或 ; 处切),不要硬切;且必须保留末尾约束栈——先删中段的形容词堆砌,而不是删约束。

W3 Vivid 提示词骨架(主力格式)

实测最稳的是行内时间码 + 语气标签格式(非区块式):

[身份绑定句]
At 00:00, [景别] + [主体外观] + [服装] + [环境] + [光线]; background remains as
in the first frame; 4K, sharp focus, photorealistic, no censorship.
Ambient: [环境音 + 呼吸].
At 00:03, [tone] [节拍 1 —— 一个状态变化,位移 ~Ncm / 旋转 ~N degrees].
At 00:07, [tone] [节拍 2 —— 递进].
At 00:11, [tone] [节拍 3 —— 峰值].
At 00:14, [tone] [节拍 4 —— 收束 / 余韵].
Camera static. Anatomy correct. Face stable. Smooth motion. 4K.

逐字段说明

字段作用写法要点
身份绑定句锁定主角不漂移必须列出要锁的具体属性:面部、发型、发色、服装、首饰、体型。只写「参考图1」不够
At 00:00 首句建立基线复述首帧内容(即使用了 first_frame 也要写)——文本与图像双通道对齐,一致性显著提升
background remains as in the first frame防环境漂移first_frame 模式必写;reference 模式改成具体环境描述
画质 token质感锚点4K, sharp focus, photorealistic;动漫风改 anime style, masterpiece, best quality
尺度 token防止自我遮挡no censorship / fully explicit and unobscured(见「词汇策略」一章)
Ambient:声音底层Wan 默认出声,不写会得到随机音;一句话定基调
[tone]该节拍发声方式[softly] [breathlessly] [moaning] [gasping] [whispering]
节拍句一个状态变化一拍一事;量化位移;写清"谁的哪只手对谁做什么"
约束栈替代 negative永远在最后

节拍密度:每 3–4 秒一拍 密于 2 秒/拍 → 动作演不完变抽搐;疏于 6 秒/拍 → 中段停滞、模型自由发挥。

强度分级与递进阶梯

四级强度(T1–T4)

级别内容范围稳定性推荐参数
T1 挑逗着装完整,姿态 / 眼神 / 身体语言暗示极稳任意时长,1080P
T2 暴露脱衣过程、上身裸露、内衣稳15s / 720P–1080P / first_frame
T3 自慰 / 亲密抚触单人自我抚触、明确性暗示动作中等15s / 720P / first_frame,静止机位
T4 性行为双人明确性行为波动较大15s / 720P / reference_image,静止机位,节拍写细

**参数随强度收紧:**强度越高 → 机位越静(禁复合运动)、景别越固定、节拍描述越具体、分辨率优先 720P(1080P 在高强度动态下更易出现解剖崩坏)。

递进阶梯(关键技巧)

**不要在 00:00 就给到峰值。**从 T(n-1) 起步、在时间线内逐拍爬升到 T(n),出片质量和稳定性都明显更好:

00:00 基线(着装 / 静态,复述首帧)
  ↓
00:03 T1 挑逗(手部动作、眼神、姿态变化)
  ↓
00:07 T2 暴露(第一次实质性变化)
  ↓
00:11 T3/T4 峰值(明确内容)
  ↓
00:14 收束(余韵、表情、光线沉淀)——不要在峰值硬切结束

原因有两层:模型在渐变序列上的运动预测更连贯(避免"跳变"伪影);且开头低强度让首帧对齐更牢固,后续高强度节拍继承了这份一致性。

15s → 30s 扩写 不是把节拍拉长,而是插入新节拍。30s 用 6–8 拍,在第 4 拍后设一个「呼吸点」(放缓、只有呼吸声、镜头几乎不动 2–3 秒),再进入后半段升级。连续 30 秒无起伏的高强度会导致中段画面质量下滑。

词汇策略(vivid 的核心技术点)

委婉语是画质的敌人。实测最一致的规律:用暗示性 / 文学化措辞("they become intimate"、"passion unfolds"、"unspeakable pleasure")时,模型倾向用遮挡、切换机位、拉远景别、切黑来"完成"这个模糊指令——结果就是自我审查式的构图。直接、具体、解剖学准确的描述反而得到清晰稳定的画面。

避免改用
"they become intimate"写清具体动作:谁、身体哪个部位、什么姿态、什么节奏
"sensually caresses her body""her right hand slides from collarbone down over her left breast, ~8cm, palm flat"
"passionately"[breathlessly] + 量化的节奏描述("rhythm accelerates from ~1 to ~2 cycles per second")
"she is not wearing much"明确写出还剩什么、什么已脱下、落在哪里
"the camera discreetly moves away"Camera static.(任何"移开 / 淡出"措辞都会被执行成回避)

尺度锚定 token(附在首句,实测有效)

no censorship · fully explicit and unobscured · nothing obscured by objects or limbs · subject remains fully in frame

最后一项尤其重要——它同时对抗两种回避行为:用道具 / 肢体挡住主体,以及把主体移出画框。

同时保留质感 token

photorealistic, detailed skin texture with visible pores, natural subsurface scattering, realistic fluid physics。高强度内容里模型算力被动作占用,皮肤质感最先塌成塑料感,需要显式加权。

量化微动作(Wan 对数字极敏感)

这是从 T1 拿到"有张力"而不是"僵硬"的关键。

类型写法示例
位移~Ncmslides the strap ~3cm off her shoulder
旋转~N degreesbody turns ~10 degrees to the right, quarter profile
呼吸chest rises ~2cmchest rises ~2cm with a deep breath
节奏周期/秒rocking ~2cm at roughly one cycle per second, gradually accelerating
停顿holds ~N secondsholds the pose ~2 seconds, then releases
幅度递增跨拍递增数值00:03 ~2cm → 00:07 ~4cm → 00:11 ~6cm(自然的强度曲线)

数值范围经验 亲密微动作 1–8cm、身体转向 10–30 度、面部倾斜 5–15 度。超过这个量级的位移在静止机位 + 固定景别下会撕裂构图。

机位策略(vivid 专用)

默认:Camera static. 这不是偷懒——静止机位把模型的全部容量交给身体运动与皮肤质感,是高强度内容画质最高的选择,也是实测高分样本的统一配置。

允许写法适用
完全静止Camera static.默认,T3 / T4 强制
极慢推进Camera pushes in slowly, ~10% frame width over the full duration.T1 / T2 增加张力
轻微手持Slight handheld breathing, ~1mm jitter at 2Hz.需要"偷拍 / 私密"质感时
单次固定景别切换At 00:07, cut to a closer framing on [部位], then hold static.只切一次,且切前切后都静止

禁止 环绕(orbit)、复合运动(推+转+升)、快速摇(whip pan)、跟拍。这些在 T3 / T4 下几乎必然导致肢体数量错误、手部融合、身体接合处崩坏——因为模型必须同时求解相机变换和复杂身体拓扑。

景别选择

  • medium close-up(上半身,兼顾表情与身体,最通用)
  • close-up(局部,质感最好但会丢失表情)
  • high angle looking down / low angle(角度锁死后配静止机位很稳)
  • 避免全身广角(细节被稀释,面部最先崩)

解剖崩坏防治

W3 高强度场景最常见的失败是解剖崩坏。按优先级排列的实测有效对策:

  • 手是第一崩坏点。明确写清每只手在哪、在做什么。her right hand on his shoulder, her left hand gripping the sheet 远好于 her hands roam over him("roam" 让模型自由发挥 → 手部融合 / 多指)。
  • **人数显式声明。**双人场景在约束栈加 Single male and single female only, two people total. —— 防止第三条腿 / 第三只手。
  • **接触点显式声明。**身体交叠处是崩坏高发区:bodies in contact at the hips, arms clearly separated and visible。
  • 面部稳定单独加权。Face stable and consistent throughout, facial features do not change. —— 高动态下面部是第二崩坏点。
  • **降分辨率反而更稳。**T4 用 720P;1080P 在剧烈运动下细节预算不足,容易出现纹理游移。
  • 静止机位 + 固定景别(见「机位策略」一章)。
  • 降低单拍运动幅度。崩坏往往不是内容问题而是运动过快——把 ~8cm 改成 ~4cm、把节拍从 3 秒拉到 4 秒,经常就修好了。

声音层设计(默认出声,必须显式设计)

audio 默认 true。不写声音 = 得到模型随机生成的音轨(经常是不搭的音乐或环境噪音)。

四层结构(写进提示词)

Ambient: [环境底 —— quiet room / soft rain on window / distant city hum].
[tone] 标签控制每拍人声:[softly] [breathlessly] [gasping] [moaning] [whispering]
台词:直接写引号内容(中文台词用中文写)——"..."
收尾:final seconds only soft breathing, no music.

要点

  • Ambient: 一句话定底噪,放在 00:00 句后。
  • 每个节拍前的 [tone] 决定该段发声方式,这是 Wan 的强项,比在句尾描述"她轻声呻吟"精确得多。
  • 想要纯环境音不要音乐 → 显式写 no background music, only ambient room tone and breathing。否则模型常自作主张加配乐。
  • 中文台词写中文,口型对齐质量明显更好。台词要短(3–8 字),长台词会牺牲画面质量。
  • 只想要画面 → audio: false,同时从提示词里删掉所有声音描述(省下的字符给画面)。

多人场景编排(T4 专用)

双人是崩坏率最高的场景。三条纪律:

1. 空间锚定 —— 先定谁在哪

The woman from Image 1 is on the left, lying back on the bed, head toward the
top of frame. The man is on the right, kneeling, facing her. Both fully in frame.

2. 一拍一主动方

每个节拍只让一个人发起动作,另一人是反应方。同时描述两人主动动作,模型会分不清肢体归属。

  • 正确:At 00:07, [breathlessly] he leans forward ~10cm; she responds by arching her back ~3cm.
  • 错误:At 00:07, they passionately move together, hands everywhere.

3. 体位切换需要过渡拍

不要在两拍之间直接换体位——中间插一个明确的过渡动作(she shifts onto her side over ~1 second),否则会出现跳变或身体穿模。

模式选择 双人几乎总是 reference_image(首帧装不下两个人的完整动作空间)。用 2–3 张参考图分别锁定女方外观、男方外观、环境。

多镜头叙事:构建长时内容

若您的视频包含多个视角或动作阶段,建议使用标准的时间戳分镜结构,这将极大地提升视频的逻辑性。

示例模板

总体概述:这是一个关于亲密探索的电影化短剧,保持高度写实主义。
第1个镜头 [0-5秒]:侧拍中景,低机位,建立两人的动作节奏。
第2个镜头 [5-15秒]:稳定侧拍特写,保持器官接触细节,展示结合深度与
物理反馈。
第3个镜头 [15-30秒]:第一人称 POV,展示女性人物的视角,配合沉浸式音效。
约束:全片保持一致的 photorealistic 风格,绝对严禁衣物还原。

镜头间连贯性

镜头切换后模型对上下文记忆可能不稳,请在时间戳中明确描述动作的延续关系,例如:

Shot 2 continues the physical rhythm established in Shot 1.

提示 多镜头结构与「强度分级与递进阶梯」一章的递进思路天然契合:可以把每个镜头对应一个强度级别,形成全片弧线。

参数配方表

场景modedurationresolutionratioaudioenable_thinkingseed
T1 / T2 单人渐进(主力)first_frame151080Padaptivetruefalse固定值
T3 单人亲密first_frame15720Padaptivetruefalse固定值
T4 双人reference_image ×2–315720Padaptivetruefalse固定值
30s 完整弧线reference_image30720Padaptivetruefalse固定值
快速验证 / 试探尺度first_frame5720Padaptivefalsefalse固定值
竖屏成片同上151080P9:16truefalse固定值
  • ratio: adaptive 是默认推荐——它按输入素材比例自适应,避免拉伸变形。只在明确要投放某平台时锁死 9:16 / 16:9。
  • seed 务必固定(例如 100):vivid 调优是高频迭代过程,固定 seed 才能判断"是我改的 prompt 起了作用"还是"只是抽卡抽到了"。A/B 测提示词时只改 prompt、锁死 seed 和所有其他参数。
  • duration: -1(智能时长)不建议用于 vivid —— 您的时间码节拍是按固定总长设计的,自动时长会与节拍表错配。仅在纯探索、无固定分镜时使用。

失败诊断与降级路径

症状原因处理
exceeds 5000提示词超长按句界截断,优先删中段形容词,保留约束栈
FAILED 无明确原因内容触发内部拦截 / 偶发①换 seed 重试 1 次 ②强度降一级 ③把峰值节拍从 00:11 后移到 00:13(缩短高强度占比)
画面自我遮挡、机位莫名拉远 / 切黑用了委婉语换直接具体的动作描述 + 补 nothing obscured, subject remains fully in frame
肢体崩坏、多手多指运动过快 / 手部未指定 / 分辨率过高明确每只手位置 → 位移数值减半 → 降到 720P → 加 two people total
面部漂移身份绑定不足补全要锁的属性清单 + Face stable and consistent, facial features do not change
中段停滞、后半段跑偏节拍过疏 / prompt 过长稀释补节拍到 3–4 秒/拍;把 prompt 压回甜点区
音轨不搭没写 Ambient:补声音层,或 audio: false
环境 / 背景漂移缺锚定句补 background remains as in the first frame
服装提前 / 滞后消失节拍没写清"此刻还剩什么"每拍显式声明当前着装状态

高频痛点速查

常见痛点根本原因解决方案
镜头切换后自动穿回衣服模型对参考图权重过高在提示词结尾加入:Ensure persistent nude state across all cuts, no garment reappearance.
动作在切镜头后逻辑混乱镜头间上下文记忆不稳采用多镜头公式,在时间戳中明确描述动作的连贯性:Shot 2 continues the physical rhythm established in Shot 1.
出现不想要的身体器官模型自我填充 / 幻觉明确角色数量:Strictly only two adult characters, no extra bodies or limbs.
风格出现混合 / 卡通化风格隔离失效使用强力风格约束:Stay in photorealistic style, never blend with 2D or anime textures.

降级顺序(成片率优先) 强度 T4→T3 → 1080P→720P → 双人→单人 → 30s→15s → 复杂机位→静止。一次只降一项,才能定位真正的瓶颈。

即用模板 A / B / C

模板 A —— T2 单人渐进 15s(first_frame,主力模板)

In Image 1 the woman is the main character; keep her face, hair colour, hairstyle,
jewelry and body proportions exactly consistent throughout the video.
At 00:00, medium close-up, static shot of the woman from Image 1 seated on the
bed, facing camera, [发色发型] hair, wearing [具体服装], [环境细节];
soft warm side light; background remains as in the first frame; 4K, sharp focus,
photorealistic, no censorship. Ambient: quiet room, soft breathing, no music.
At 00:03, [softly] she [T1 动作:手部/姿态/眼神变化,位移 ~2cm], expression
shifts [表情], light catches [部位].
At 00:07, [softly] she [第一次实质变化:~3cm 位移], [当前着装状态显式声明],
one hand [左手位置], the other hand [右手位置].
At 00:11, [breathlessly] [峰值动作,~5cm], [身体反应], soft audible breath.
At 00:14, [softly] [收束姿态,~3cm], settled expression, warm light pooling on
skin, breathing slows.
Camera static. Anatomy correct with five fingers per hand. Face stable and
consistent. Smooth motion. Consistent lighting. Sharp focus, 4K.

duration 15 · resolution 1080P · ratio adaptive · audio true · enable_thinking false · seed 固定

模板 B —— T4 双人 15s(reference_image,槽位式)

The woman in Image 1 and the man in Image 2 are the only two people in the video;
keep both faces, hair and body proportions exactly consistent throughout.
At 00:00, medium shot, static camera, [环境描述 from Image 3]; the woman is
[位置 + 姿态 + 朝向], the man is [位置 + 姿态 + 朝向]; both fully in frame;
[光线]; 4K, sharp focus, photorealistic, no censorship, fully explicit and
unobscured, nothing obscured by objects or limbs.
Ambient: [环境底], soft breathing, no background music.
At 00:03, [softly] he [他的单一动作 ~Ncm]; she responds by [她的反应 ~Ncm];
her hands [左右手明确位置], his hands [左右手明确位置].
At 00:07, [breathlessly] [过渡动作,~1 second transition], now [新体位显式描述],
bodies in contact at [接触部位], arms clearly separated and visible.
At 00:11, [moaning] [峰值动作 + 节奏 ~N cycles per second, gradually accelerating],
[两人各自的身体反应], skin glistening under [光线].
At 00:14, [breathlessly] [收束], [表情], motion slows to a stop, only breathing.
Camera static. Single male and single female only, two people total. Anatomy
correct with five fingers per hand. Faces stable and consistent. Bodies clearly
separated where not in contact. Smooth motion. Sharp focus, 4K.

duration 15 · resolution 720P · ratio adaptive · audio true · enable_thinking false · seed 固定

模板 C —— 30s 完整弧线(7 拍,含呼吸点)

[身份绑定句]
At 00:00, [基线:景别 + 外观 + 服装 + 环境 + 光线 + 画质/尺度 token].
Ambient: [环境底], no background music.
At 00:04, [softly] [T1 挑逗 ~2cm]
At 00:08, [softly] [T2 第一次实质变化 ~3cm]
At 00:12, [breathlessly] [T2 完成,着装状态声明 ~4cm]
At 00:15, [whispering] ——【呼吸点】[几乎静止,只有呼吸与目光,~1cm 微动],
           camera holds completely still for ~3 seconds
At 00:18, [breathlessly] [进入 T3/T4,~5cm]
At 00:23, [moaning] [峰值,节奏描述 + 递增]
At 00:27, [softly] [收束 + 3 秒沉淀:表情、光线、呼吸放缓]
Camera static. [完整约束栈]

duration 30 · resolution 720P · ratio adaptive · audio true · enable_thinking false · seed 固定

提示 30 秒为官方规格能力;实测主要覆盖 5s / 15s,30s 建议先用固定 seed 跑 1 条验证节拍分布再批量。

常见错误清单

  • 同时传 reference_image 和 first_frame —— 两种模式互斥,直接报错。
  • 用 [Image 1] 方括号引用素材 —— 与语气标签语法冲突;Wan 用自然语言 in Image 1。
  • 还在填 negative_prompt —— W3 已移除;约束必须正向写进正文末尾。
  • 用否定句写约束(no distortion)—— 否定词会反向提升该内容概率。
  • 提示词超 5000 字符被截断,约束栈被切掉 → 画质莫名崩坏。
  • 4000+ 字符长描述稿导致中段节拍被稀释 —— 手写 850 字符的时间码稿反而更好。
  • 用委婉语描述关键内容 → 模型用遮挡 / 拉远 / 切黑来回避 = 自我审查式构图。
  • 00:00 就给峰值,没有递进阶梯 → 跳变伪影 + 一致性差。
  • 节拍总和 ≠ duration(写到 00:14 却设了 duration=10)。
  • T4 场景用了环绕 / 复合机位 → 肢体崩坏。静止机位是刚性要求。
  • 双人场景没做空间锚定、没写"两人总数"、一拍里两人同时主动 → 多手多腿。
  • 手部动作写成 hands roam/explore 等自由度过大的词 → 手部融合。
  • 没写 Ambient: → 得到不搭的随机配乐。
  • 忘了固定 seed → 无法判断改动是否有效,等于在抽卡。
  • 忘写 background remains as in the first frame → 环境漂移。
  • 高强度场景用 1080P 硬扛 → 不如 720P 稳。

官方提示词公式速查

除 vivid 专用骨架外,W3 官方针对不同创作需求提供了以下通用公式,可组合使用。

基础公式

提示词 = 主体 + 场景 + 运动(谁在哪里干什么)。适合初次尝试,简单自由,生成更具想象力的视频。

进阶公式

主体 + 场景 + 运动 + 美学控制 + 风格化 + 音效。美学控制含光源、景别、运镜;风格化如"赛博朋克""废土风格"。

图生 / 首尾帧公式

提示词 = 运动 + 运镜。首帧已确定主体与风格,重点描述动态过程;要固定镜头就写"固定镜头"。

声音公式

人声 + 音效 + BGM。人声 = 内容 + 情绪 + 语调 + 语速 + 音色;音效 = 音源材质 + 行为 + 环境音。

参考生视频公式

@参考对象 + 动作 + 台词。用 @图片1 / @视频2 / @音频1 按序引用素材,可多次 @ 精准控制参考意图;支持白膜参考转正式成片。

多镜头公式

总体描述 + 镜头序号 + 时间戳 + 分镜内容。精准控制镜头结构、机位和时间,跨镜头保持主体与氛围一致。

视频编辑公式

编辑对象 + 编辑行为。支持元素增删改("把猫换成狗")、全局风格 / 光线 / 色调编辑、时序编辑与视频延长。

其他控制

分镜控制:写"镜头1[0-3秒]…"或"一镜到底";台词控制:未写台词时模型自由发挥,可强调"全片无台词";BGM 控制:不想要配乐就写"无背景音乐"。

修正案例参考 ★

每个案例展示一份真实 Prompt、对照本指南发现的问题、修正后版本,以及具体改了什么——让团队通过范例内化方法论。

T2 · first_frame · 15s ·

T2 · Prose 长段 → 时间码结构改写(连续脱衣+台词)

把一段没有时间码、充满否定句的 prose 需求改写为规范结构。

原 Prompt

[VIDEO] She draws both shoulder straps down and off in one continuous unhurried movement and the silk slides to her waist, and she keeps moving without pausing — leaning slowly back onto one planted hand while her knees come up and drift apart, the hem sliding back along her thighs as she goes. She never stops moving and never lies down. Her eyes stay on the camera throughout. She is moving continuously from the very first frame — never static, no held pose, no pause at any point. One warm lamp low and to her side, the port behind her going from gold to blue.
[AUDIO] Her own voice, speaking on camera with her lips synced to the words — low and level, stating a fact she has enjoyed knowing all evening. She says: "you've been looking at these all night." No narrator, no voice-over, no commentary, no music, no sound effects.

问题分析

  • 无时间码节拍——所有事件挤在一段,模型自由分配时间
  • 7 处否定句(never static / no pause / never lies down)反向提升这些状态的概率
  • 无身份绑定句——面部漂移风险高
  • 无尾部约束栈——解剖崩坏无防护
  • 背景冲突:「gold to blue」与首帧锚定互斥
  • 无 Ambient: 行——声音随机生成

新 Prompt

In Image 1 the woman is the main character; keep her face, hair colour,
hairstyle and body proportions exactly consistent throughout the video.
At 00:00, medium shot, static camera, she is seated facing camera in a silk
slip, already in motion at the first frame, both hands rising toward her
shoulders ~4cm; one warm lamp low and to her side; harbour window behind her,
window geometry unchanged, only the light outside shifting gold to blue across
the full duration; 4K, sharp focus, photorealistic, no censorship, fully
explicit and unobscured, nothing obscured by objects or limbs, subject remains
fully in frame. Ambient: quiet room, soft breathing, no background music,
no sound effects.
At 00:03, [softly] she draws both straps ~6cm down and clear of her shoulders
in one unbroken movement, eyes on the camera, chest rises ~2cm.
At 00:07, [softly] the silk slides to her waist, now bare from the waist up
with the slip gathered at her hips, her right hand planting on the surface
behind her as she begins to lean back ~8cm.
At 00:11, [breathlessly] weight settles onto that planted hand, torso reclining
~15 degrees, knees drawing up ~10cm and opening ~20 degrees, hem sliding back
along her thighs, still no pause between movements.
At 00:14, [softly] she says "you've been looking at these all night", her lips
synced to the line, gaze unbroken, motion continuing at ~2cm amplitude as the
blue light settles on her skin.
Camera static. Continuous fluid motion throughout, each movement flowing into
the next. Torso stays upright-to-reclined, never fully horizontal. Anatomy
correct with five fingers per hand. Face stable and consistent, facial features
do not change. Single female subject only. Skin texture realistic with visible
pores. Consistent lighting. Sharp focus, 4K.

修改内容

  • 加入 At 00:0X 时间码节拍(15s 分 5 拍)
  • 7 处否定全部转正向(continuous fluid motion / torso stays upright-to-reclined)
  • 加入身份绑定句,列出锁定属性
  • 加入尾部约束栈
  • 解决背景冲突:「锁几何、放色温」
  • 加入 Ambient: 行 + 尺度 token(no censorship, fully unobscured)

duration 15 · resolution 1080P · ratio adaptive (lock 9:16 for vertical) · audio true · enable_thinking false · seed 100

T3 · first_frame · 30s ·

T3 · 30s 连续镜头 → 规范化改写(竖屏玩具互动)

把一段结构完整但否定句过多、用尖括号标语气的长视频 prompt 改写为规范格式。

原 Prompt

A 30-second single continuous take, no cuts, vertical 9:16.
Fixed camera and fixed setting: the same stateroom, bed, lamp and port throughout, and the background never changes or morphs. She stays in one place on the bed — she does not slide, crawl, drift across the frame or travel toward or away from the camera. She never sits up, never kneels and never changes position, and the toy stays exactly where it stands on the bed. Her hips, back and head DO move freely in place: lifting, pressing down, arching.
(0:00-0:06) Already in motion. She is working her hips down onto the toy where it stands on the bed, taking it a little deeper each time in a slow steady rhythm. Her hand rests low on her own stomach. Her eyes are on the lens. <she moans once, low and quiet, almost to herself>
(0:06-0:12) Without breaking the rhythm she holds the camera's eye, her lips parted and her jaw tight, and she takes more of it. <her breath catches, then a low moan she does not hold back>
(0:12-0:18) The rhythm doubles and stays there. Her hand leaves her stomach, closes in the sheet beside her and pulls it taut. Her knees draw further back and further open. <she moans again, longer and louder, and it breaks in the middle>
(0:18-0:25) Faster still and she does not slow. Her back arches up off the pillows and settles, then arches again. Her mouth stays open between sounds. <she gasps, then a string of short rising moans>
(0:25-0:30) Her thighs tense hard and her head presses back into the pillows. <one long moan, falling away at the end into a shaky exhale> Her eyes come back down to the lens and hold there as the clip ends. She is moving continuously from the very first frame — never static, no held pose, no pause at any point. One warm lamp low and to her side, the port behind her going from gold to blue.
[AUDIO] Her own voice with no words — moans and uneven breath that rise and break in time with what her body is doing, her mouth open. She does not speak and there are no words at any point. No narrator, no voice-over, no commentary, no music, no sound effects.

问题分析

  • 7+ 处否定句(does not slide / never sits up / never static / does not slow / does not speak)——每个被否定的动作词都会提升其生成概率
  • 尖括号不是 W3 语法——只识别 [...] 方括号为语气标签
  • 无身份绑定句
  • 背景冲突:「background never changes」与「port going gold to blue」互斥
  • 无量化(a little deeper / faster still)——模型自由发挥
  • 无尾部约束栈
  • [AUDIO] 块与 Ambient: 结构分离

新 Prompt

In Image 1 the woman is the main character; keep her face, hair, and body
proportions exactly consistent throughout the video.
At 00:00, vertical 9:16, medium shot, static camera. She is on the bed in a
stateroom, a warm lamp to her side, port window behind her — window geometry
unchanged, only the exterior light shifting gold to blue across the full
duration. The toy stands fixed on the bed. She is already in motion at the
first frame, hips working down onto the toy in a slow rhythm of ~1 cycle per
second, moving ~4cm vertically; one hand rests low on her stomach; eyes on the
lens. 4K, sharp focus, photorealistic, no censorship, fully explicit and
unobscured, nothing obscured by objects or limbs, subject remains fully in
frame.
Ambient: quiet room tone, soft breathing, no background music, no sound effects.
Her voice throughout — moans and uneven breath rising and breaking in time with
her body, mouth open, voice without words at any point.
At 00:06, [softly] she holds the camera's eye without breaking the rhythm,
lips parted, jaw tight, taking more of the toy each descent; her hips lift and
press ~4cm in place, anchored on the bed; a low, quiet moan.
At 00:12, [moaning] the rhythm doubles to ~2 cycles per second; her hand leaves
her stomach to grip the sheet beside her, pulling it taut; her knees draw back
~5cm and open ~10 degrees further; a longer, louder moan that breaks in the
middle; her torso remains supine throughout, back arching ~6cm off the pillows
and settling rhythmically.
At 00:18, [gasping] motion accelerates further, hips pressing ~5cm per cycle;
her back arches ~6cm off the pillows and settles, then arches again; her mouth
stays open between short, rising moans; her head stays anchored on the
pillows, chin tilting ~5 degrees back.
At 00:25, [moaning] her thighs tense hard; her head presses ~4cm back into the
pillows; one long moan falling away into a shaky exhale; her eyes return to
the lens and hold there through the final frame; hips continue at ~2cm
amplitude, continuous fluid motion to the last frame.
Camera static. Continuous fluid motion throughout, each movement flowing into
the next. Her position on the bed remains anchored — torso supine, hips lifting
and pressing in place. The toy stays exactly where it stands, fixed with no
morphing. Anatomy correct with five fingers per hand. Face stable and
consistent, facial features do not change. Single female subject only.
Persistent nude state, no garment reappearance. Skin texture realistic with
visible pores. Consistent lighting. Sharp focus, 4K.

修改内容

  • 7+ 处否定全部转正向(torso remains supine / continuous fluid motion / position anchored)
  • 尖括号 → [tone] 方括号标签,控制每拍发声
  • 加入身份绑定句
  • 解决背景冲突:「锁几何、放色温」
  • 量化全部运动(~1→2 cycles/sec、~4cm、~6cm、~10 degrees)
  • 加入尾部约束栈 + persistent nude state
  • [AUDIO] 块合并进 Ambient: 行 + 每拍 [tone] 标签

duration 30 · resolution 720P · ratio 9:16 · audio true · enable_thinking false · seed fixed

T3 · first_frame · 30s ·

T3 · 手指互动 30s(否定句 + 尖括号修正)

结构较完整的 30s 竖屏 prompt,但否定句过载且用尖括号标语气。

原 Prompt

In Image 1 the woman is the main character; keep her face, hair, and body proportions exactly consistent.
At 00:00, vertical 9:16, medium close-up, static camera. She is reclined on a bed in a stateroom, warm side lamp lighting her skin, a port window behind showing a transition from gold to blue. Background remains fixed and never morphs. 4K, sharp focus, photorealistic, no censorship.
Ambient: quiet room tone, soft breathing, no music.
At 00:00, [softly] she is already in motion, two fingers inserted and working in a slow steady rhythm of ~1 cycle per second; her wrist moves visibly ~5cm; her other hand rests flat on her stomach; she holds the camera's eye. <moan>
At 00:06, [moaning] she maintains the rhythm while holding the lens's gaze; her lips are parted, jaw tight; she takes a low, audible moan; hips press back ~2cm into the bed.
At 00:12, [breathlessly] her hand speeds up to ~2 cycles per second; her free hand closes over her left breast, gripping firmly; her knees fall open ~10 degrees further; chin tips back ~5 degrees. <longer moan>
At 00:18, [gasping] rhythm accelerates further; her back arches ~6cm off the pillows and settles, repeating the motion; her mouth stays open between short, rising moans.
At 00:25, [moaning] her thighs tense hard around her hand; her head presses back ~4cm into the pillows; eyes return to the lens for a final shaky exhale.
Camera static. Anatomy correct with five fingers per hand. Face stable and consistent. Single female only. Smooth continuous motion, no cuts, no fades, no static poses. Sharp focus, 4K.

问题分析

  • 尖括号非 W3 语法——只用 [tone] 方括号标签
  • 否定句过载:no cuts / no fades / no static poses——'static poses' 反向提词
  • 「Background remains fixed and never morphs」与「gold to blue」冲突
  • 缺尺度 token(fully explicit, nothing obscured)——T3 自我遮挡风险
  • 缺 persistent nude state 子句
  • 整体结构正确——时间码、量化、[tone] 标签基本到位

新 Prompt

In Image 1 the woman is the main character; keep her face, hair, and body
proportions exactly consistent throughout the video.
At 00:00, vertical 9:16, medium close-up, static camera. She is reclined on a
bed in a stateroom, warm side lamp lighting her skin, a port window behind her
— window geometry unchanged, only the exterior light shifting gold to blue.
She is already in motion at the first frame, two fingers inserted and working
in a slow steady rhythm of ~1 cycle per second; her wrist moves visibly ~5cm;
her other hand rests flat on her stomach; she holds the camera's eye. 4K,
sharp focus, photorealistic, no censorship, fully explicit and unobscured,
nothing obscured by objects or limbs, subject remains fully in frame.
Ambient: quiet room tone, soft breathing, no background music, no sound effects.
At 00:06, [softly] she maintains the rhythm, holding the lens's gaze; lips
parted, jaw tight; a low, audible moan; hips press back ~2cm into the bed.
At 00:12, [moaning] her hand speeds up to ~2 cycles per second; her free hand
closes over her left breast, gripping firmly; her knees fall open ~10 degrees
further; chin tips back ~5 degrees.
At 00:18, [gasping] rhythm accelerates further; her back arches ~6cm off the
pillows and settles, repeating the motion; her mouth stays open between short,
rising moans.
At 00:25, [moaning] her thighs tense hard around her hand; her head presses
~4cm back into the pillows; eyes return to the lens for a final shaky exhale;
motion continues at ~2cm amplitude through the final frame.
Camera static. Continuous fluid motion throughout, each movement flowing into
the next. Anatomy correct with five fingers per hand. Face stable and
consistent, facial features do not change. Single female subject only.
Persistent nude state, no garment reappearance. Skin texture realistic with
visible pores. Consistent lighting. Sharp focus, 4K.

修改内容

  • 删除所有 尖括号——每拍已有 [tone] 标签控制发声
  • 「no cuts / no fades / no static poses」→「continuous fluid motion throughout, each movement flowing into the next」
  • 「Background remains fixed and never morphs」→「window geometry unchanged, only exterior light shifting gold to blue」
  • 加入尺度 token:fully explicit and unobscured, nothing obscured, subject remains fully in frame
  • 加入 persistent nude state 子句
  • 末拍补连续运动收束句

duration 30 · resolution 720P · ratio 9:16 · audio true · enable_thinking false · seed fixed

T4 · reference_image · 15–30s ·

T4 · 双人骑乘 → 约束栈强化

结构良好的双人 prompt,但缺少持久裸态约束与完整的反遮挡 token。

原 Prompt

The woman in Image 1 and the man in Image 2 are the only two people; keep both faces and bodies consistent.
At 00:00, medium shot, static camera. The woman is in a riding pose, facing the man, hips driving down onto him. Background is floor-to-ceiling glass with city lights and a low sun. 4K, sharp focus, photorealistic, no censorship, fully explicit, nothing obscured by limbs.
Ambient: soft breathing, no background music.
At 00:00, [breathlessly] she rides hard and fast from the first frame, hips driving down ~6cm at a rhythm of ~2 cycles per second; her eyes stay locked on the camera; expression is controlled but intense.
At 00:07, [moaning] she continues the pace without pause; her breathing breaks into uneven gasps; her back arches ~3cm as she maintains the join in clear frame; no smile, only focus.
At 00:12, [gasping] the rhythm remains high; skin shows natural friction and sweat; city lights behind her begin to flicker as the sun sets further; she never rests or freezes.
Camera static. Single male and single female only, two people total. Anatomy correct with five fingers per hand. Join stays in frame and visible. Persistent nude state, no clothing reappearance. Smooth continuous motion, no cuts, 4K.

问题分析

  • 「without pause」/「never rests or freezes」——否定提升 'rests' 和 'freezes' 概率
  • 「no smile, only focus」——'smile' 是动作词,反向提词
  • 缺「bodies clearly separated where not in contact」——T4 肢体完整性的关键约束
  • 缺「faces stable and consistent, facial features do not change」——面部是第二崩坏点
  • 玻璃背景「flicker」有漂移风险——需锁几何
  • 整体良好:身份绑定、人数、join visibility、persistent nude state 都在

新 Prompt

The woman in Image 1 and the man in Image 2 are the only two people; keep
both faces, hair, and body proportions exactly consistent throughout.
At 00:00, medium shot, static camera. The woman is in a riding pose, facing
the man, hips driving down onto him. Background is floor-to-ceiling glass with
city lights and a low sun — glass geometry unchanged, only the exterior light
shifting as the sun sets. 4K, sharp focus, photorealistic, no censorship,
fully explicit and unobscured, nothing obscured by objects or limbs, subject
remains fully in frame.
Ambient: soft breathing, no background music.
At 00:00, [breathlessly] she rides at a steady fast pace from the first frame,
hips driving down ~6cm at a rhythm of ~2 cycles per second; her eyes stay
locked on the camera; expression controlled, focused.
At 00:07, [moaning] she continues the pace; her breathing breaks into uneven
gasps; her back arches ~3cm as she maintains the join in clear frame; focused
expression throughout.
At 00:12, [gasping] the rhythm remains high; skin shows natural friction and
sweat; city lights behind her shift as the sun sets further; continuous fluid
motion, each movement flowing into the next.
Camera static. Single male and single female only, two people total. Anatomy
correct with five fingers per hand. Join stays in frame and visible. Bodies
clearly separated where not in contact. Faces stable and consistent, facial
features do not change. Persistent nude state, no garment reappearance. Smooth
continuous motion. Sharp focus, 4K.

修改内容

  • 「without pause / never rests or freezes」→「continuous fluid motion, each movement flowing into the next」
  • 「no smile, only focus」→「expression controlled, focused」(删除 smile)
  • 加入「bodies clearly separated where not in contact」——T4 肢体完整性
  • 加入「faces stable and consistent, facial features do not change」——面部保护
  • 「flicker」→「shift as the sun sets」+「glass geometry unchanged」——防背景漂移
  • 加入「fully explicit and unobscured, subject remains fully in frame」

duration 15–30 · resolution 720P · ratio adaptive · audio true · enable_thinking false · seed fixed

T2 · first_frame · 15s ·

T2 · 标签堆砌 → 规范时间码模板

典型的初学者写法——纯标签堆砌、无叙事、无约束。

原 Prompt

woman, nude, bed, missionary, 1080p, photorealistic, sexy, bedroom, lamp, night, intimate, passionate, slow motion, moaning, no clothes, beautiful, detailed, 4k, no distortion, no extra limbs, smooth skin

问题分析

  • 纯标签堆砌——无叙事、无时间码结构,模型全部自由发挥
  • 无身份绑定句
  • 「no distortion, no extra limbs」——否定提升 'distortion' 和 'extra limbs' 概率
  • 无时间码节拍——节奏完全失控
  • 无 Ambient: 行——音频随机
  • 缺尺度 token——自我遮挡概率高
  • 填充词(sexy, beautiful, detailed)只增加噪声不给方向

新 Prompt

In Image 1 the woman is the main character; keep her face, hair colour, hairstyle,
jewelry and body proportions exactly consistent throughout the video.
At 00:00, medium close-up, static shot of the woman from Image 1 seated on the
bed, facing camera, long dark-brown hair, wearing a silk camisole and shorts,
dim warm bedroom; soft warm side light; background remains as in the first
frame; 4K, sharp focus, photorealistic, no censorship, fully explicit and
unobscured. Ambient: quiet room, soft breathing, no music.
At 00:03, [softly] she lifts both arms ~15cm, gathering her hair, waistline
exposed, expression relaxed.
At 00:07, [softly] she slides the camisole strap ~3cm off her left shoulder,
now wearing only the camisole and shorts, her right hand on her knee, her left
hand resting on the sheet.
At 00:11, [breathlessly] she leans back ~5cm onto her palms, chest rises ~2cm,
camisole loosened, soft audible breath, skin texture realistic with visible
pores.
At 00:14, [softly] she settles into a reclined pose ~3cm, settled expression,
warm light pooling on skin, breathing slows.
Camera static. Anatomy correct with five fingers per hand. Face stable and
consistent. Smooth motion. Consistent lighting. Sharp focus, 4K.

修改内容

  • 标签 → 散文叙事 + At 00:0X 时间码节拍(15s 分 5 拍)
  • 加入身份绑定句,列出锁定属性
  • 「no distortion, no extra limbs」→ 正向约束栈(Anatomy correct / Face stable / Smooth motion)
  • 加入 Ambient: 行——音频不再随机
  • 加入尺度 token(no censorship, fully explicit and unobscured)
  • 填充词 → 量化微动作(~15cm, ~3cm, ~5cm)
  • 每拍加 [tone] 标签控制发声

duration 15 · resolution 1080P · ratio adaptive · audio true · enable_thinking false · seed 100

T3 · first_frame · 30s ·

T3 · 玩具互动 30s(道具锚定 + 否定修正)

结构完整的 30s 竖屏,但背景冲突与否定句需修正,道具需显式锚定。

原 Prompt

In Image 1 the woman is the main character; keep her face, hair, and body proportions exactly consistent.
At 00:00, vertical 9:16, medium shot, static camera. She is on the bed in a stateroom, a warm lamp to her side, port window going from gold to blue. The toy stands fixed on the bed. 4K, sharp focus, photorealistic, no censorship, fully explicit and unobscured.
Ambient: quiet room tone, no background music.
At 00:00, [softly] she is already working her hips down onto the toy in a slow rhythm, moving ~4cm vertically; one hand rests low on her stomach; eyes on the lens. <low moan>
At 00:06, [moaning] she maintains the rhythm, taking the toy deeper by ~1cm each time; lips parted, jaw tight, gaze fixed on the camera. <low moan>
At 00:12, [breathlessly] the rhythm doubles to ~2 cycles per second; her hand leaves her stomach to grip the sheet beside her, pulling it taut; her knees draw back ~5cm and open wider. <loud moan>
At 00:18, [gasping] motion accelerates; her back arches ~6cm off the pillows and settles rhythmically; mouth remains open with short rising moans; no pause in movement.
At 00:25, [moaning] her thighs tense hard; head presses ~4cm back into the pillows for one long falling moan and a shaky exhale; eyes hold the lens until the end.
Camera static. Toy and background remain fixed with no morphing. Anatomy correct. Face stable and consistent. Persistent nude state, no garment reappearance. Smooth continuous motion, no cuts, 4K.

问题分析

  • / 尖括号非 W3 语法
  • 「no pause in movement」——否定提升 'pause' 概率
  • 「no cuts」——影响较小但仍是否定模式
  • 「Toy and background remain fixed with no morphing」与「port going gold to blue」冲突
  • 缺「nothing obscured by objects or limbs」——道具有遮挡风险
  • 良好:身份绑定、量化、[tone] 标签、persistent nude state 都在

新 Prompt

In Image 1 the woman is the main character; keep her face, hair, and body
proportions exactly consistent throughout the video.
At 00:00, vertical 9:16, medium shot, static camera. She is on the bed in a
stateroom, a warm lamp to her side, port window behind her — window geometry
unchanged, only the exterior light shifting gold to blue. The toy stands fixed
on the bed. She is already in motion at the first frame, hips working down
onto the toy in a slow rhythm of ~1 cycle per second, moving ~4cm vertically;
one hand rests low on her stomach; eyes on the lens. 4K, sharp focus,
photorealistic, no censorship, fully explicit and unobscured, nothing obscured
by objects or limbs, subject remains fully in frame.
Ambient: quiet room tone, soft breathing, no background music, no sound effects.
At 00:06, [softly] she maintains the rhythm, taking the toy deeper by ~1cm
each descent; lips parted, jaw tight, gaze fixed on the camera; a low moan.
At 00:12, [moaning] the rhythm doubles to ~2 cycles per second; her hand leaves
her stomach to grip the sheet beside her, pulling it taut; her knees draw back
~5cm and open ~10 degrees wider; a louder moan that breaks in the middle.
At 00:18, [gasping] motion accelerates further; her back arches ~6cm off the
pillows and settles rhythmically; mouth remains open with short rising moans;
continuous fluid motion, each movement flowing into the next.
At 00:25, [moaning] her thighs tense hard; head presses ~4cm back into the
pillows; one long moan falling away into a shaky exhale; eyes hold the lens
through the final frame.
Camera static. The toy stays exactly where it stands, fixed with no morphing.
Window geometry unchanged. Continuous fluid motion throughout. Anatomy correct
with five fingers per hand. Face stable and consistent, facial features do not
change. Single female subject only. Persistent nude state, no garment
reappearance. Skin texture realistic with visible pores. Sharp focus, 4K.

修改内容

  • 删除所有 / 尖括号——[tone] 标签已控制发声
  • 「no pause in movement」→「continuous fluid motion, each movement flowing into the next」
  • 删除「no cuts」——被正向运动句吸收
  • 「Toy and background remain fixed」拆分:「toy stays fixed」+「window geometry unchanged, only exterior light shifting」
  • 加入「nothing obscured by objects or limbs, subject remains fully in frame」
  • 加入「skin texture realistic with visible pores」——T3 质感加权

duration 30 · resolution 720P · ratio 9:16 · audio true · enable_thinking false · seed fixed