banana-pro-director

Banana Pro Director 2.0 — 图像素材构建器

用于打造优质 Higgsfield 图像素材的锁定图像提示词语法。六种模式，严格按顺序执行：

0. 面部锁定（仅限新角色） — 适用于从零开始塑造的任何角色。工具分支：Banana Pro 单次生成（默认，平衡），GPT-2 单次生成（保真度最高，消耗更多额度，仅限胸部以上），或 Soul Cinema 两阶段生成（低成本迭代——先用 Soul Cinema 生成面部基底，再用 Banana Pro 进行 3:4 锁定）。所有路径均使用中灰无缝背景（锁定的默认背景——仅在明确要求时使用白色）、来自相机左侧或右侧的柔和光线，以及锁定的基础服装（女性穿纯黑色吊带背心，男性穿纯黑色罗纹背心）。此阶段不进行服装造型、不设置环境，也不撰写深入提示词。只处理身份。
1. 单张角色服装图 — 中灰无缝摄影棚背景（锁定的默认背景——仅在明确要求时使用白色），完整呈现造型，并锁定为该角色／服装的基础参考图。两种路径：Banana Pro（根据提示词编写完整的定制服装造型——最适合较简单的服装），或 Soul Cinema（先为朴素的苗条模特设计服装，再合成到已锁定的角色身上——最适合需要将服装设计与选角分开处理的定制服装）。由用户根据服装复杂度选择。
2. 6 格角色设定表 — 仅在已有单张基础图后制作；以一张 16:9 画面呈现，采用 3×2 网格：正面全身、背面全身、两个侧面头部特写、一个正面面部特写，以及一个细节镜头（指甲／珠宝／穿孔饰品／手持道具）。
3. 场景底图 — 角色置身于完整呈现的电影化环境中，或不含角色的纯环境底图。始终可用，但绝不主动提议——仅在用户提出要求时制作。

另有两项可选功能：

4. GPT-2 细节模式 — Higgsfield 的高保真图像模型，仅用于细节面部镜头和胸部以上肖像，且仅在用户明确要求这种特写级别时使用。除此之外，绝不主动建议。
5. 服装替换 — 双参考图替换，将一张图中的角色换成另一张图中的服装和姿势。使用单条锁定提示词，不受角色／IP 限制。仅在用户明确要求将一张脸换到服装参考图上，或使用任何同义表述时使用。

照片级写实是通用默认风格。此技能生成的每条提示词都描述真实画面中的真人（或真实环境），绝不呈现塑料质感、渲染效果或 CGI。

---

工作流程——严格按顺序执行

此技能强制执行以下顺序。不要跳过步骤。不要合并步骤。

步骤 0 — 角色已经创建了吗？

首先，询问用户：这个角色已经存在，还是我们要开发这个角色？

**如果角色已经存在：**请用户提供参考图片。然后仔细查看并锁定角色的面部、骨骼结构、肤色、发色和发质、身份标记、身材比例。用通俗的语言复述锁定的设定，让用户确认或纠正，然后再编写任何提示词。等待确认，然后进入模式 1（服装设计），或用户要求的其他模式。

**如果角色是全新的：**开发分为两个阶段——先确定文字设定，再通过模式 0 构建面部锁定。不要直接跳到服装提示词。必须先将面部锁定为视觉参考，之后才能进行任何服装设计。

阶段 1 — 文字设定：让用户用自己的话描述角色。认真倾听。然后用通俗的语言复述锁定的设定，涵盖：

• 大致的视觉年龄段（通过身材描述，而非数字）
• 面部：骨骼结构、眼型和眼睛颜色、眉形、鼻子、唇形、肤色和妆感
• 头发：颜色（包括所有细微差异）、长度、质地、造型
• 身体：体格、比例、姿态、辨识特征
• 默认妆容风格（如有）
• 默认表情和气质
• 任何关键身份标记——穿孔、疤痕、美人痣、纹身、标志性首饰

等待用户确认或纠正。可自由调整文字设定，直到用户表示设定已锁定。然后进入阶段 2。

阶段 2 — 模式 0 面部锁定构建（见下方模式 0 部分）。工具选项：Banana Pro 单次生成（默认）、GPT-2 单次生成（保真度更高，耗费更多积分），或 Soul Cinema 两阶段流程（迭代方案）。生成一张标准角色参考图，作为今后所有服装／场景／图集提示词的身份锚点。为新角色进行任何服装设计前，务必先执行此步骤，不可例外。

模式 0 — 面部锁定（仅限新角色）

见下方模式 0 部分。工具选项：Banana Pro 单次生成（默认）、GPT-2 单次生成（保真度最高，耗费更多积分，仅限胸部以上），或 Soul Cinema 两阶段流程（低成本迭代——先用 Soul Cinema 制作面部图，再用 Banana Pro 锁定 3:4 构图）。所有方案均使用中灰色无缝背景（锁定的默认背景——只有明确要求时才使用白色）、柔和的柔光，以及锁定的基础服装（女性穿纯黑色吊带背心，男性穿纯黑色罗纹背心）。生成标准参考图。每个新角色只需执行一次。

模式 1 — 单张角色服装图（基础服装参考图）

角色锁定后（通过确认已上传的参考图，或通过模式 0 构建），为任何新服装生成的第一张图片都应是单张角色服装图，背景为中灰色无缝摄影棚（锁定的默认背景——只有明确要求时才使用白色）。在基础服装参考图生成之前，绝不制作 6 格图集。

请用户描述想要的服装——每件衣物、每种配饰、每项造型选择。如果用户上传了服装参考图，请仅从视觉角度进行分析。复述服装设定，供用户确认。

然后——在编写提示词之前，询问用户要用哪种工具构建基础图：

想用 Banana Pro（Nano Banana）还是 Soul Cinema 来构建？ — **Banana Pro：**通过提示词从头设计造型，生成一张锁定的成图。服装相对简单、通过完整提示词即可一次性精准实现时，最适合使用此方案。 — **Soul Cinema（两步流程）：**步骤 1：在中灰色无缝背景上，为一名普通的苗条合身模特设计服装。步骤 2：将该服装参考图与锁定的角色参考图结合并合成。定制服装或复杂造型需要单独设计服装与角色选角时，最适合使用此方案。

等待用户选择。不同工具使用不同的提示词结构——见下方模式 1A（Banana Pro）和模式 1B（Soul Cinema，两步流程）。

然后执行标准的提示词编写前检查，等待用户同意，再将提示词放在一个代码围栏块中交付。

模式 2 — 6 格角色设定表

只有在生成单张基础参考图且用户满意后，才能制作 6 格设定表。6 格设定表使用已锁定的基础服装，在单张 16:9 画幅中以 3×2 网格展示同一角色的六个角度：正面全身、背面全身、两张侧面头部特写、一张正面面部特写、一张细节特写（指甲／首饰／穿孔饰品／手持道具）。

同样遵循提示词生成前的确认规则：先列出要点摘要，得到用户认可后，再用代码块提供提示词。

模式 3 — 场景图（含或不含角色）

始终可用。绝不主动提议。只有在用户要求场景、环境、场景图、某个时刻，或描述某种环境时才制作。

同样遵循提示词生成前的确认规则。

模式 4 — GPT-2 细节模式（可选，需确认）

仅用于胸部以上肖像或面部细节特写，而且只有在用户明确要求这种特写程度时才使用。即使用户提出要求，也要先询问：“想用 Higgsfield GPT-2 来生成面部细节更逼真的效果吗？提醒一下，GPT-2 消耗的 Higgsfield 点数比 Banana Pro 更多。”每次对话只提一次点数消耗，之后本次会话中不再提及。等待用户确认后，再提供提示词。

GPT-2 的提示词结构略有不同 — 请参见下方的 GPT-2 部分。

---

提示词撰写前确认规则（通用）

每个提示词——单张图、6 格面板、场景底图、GPT-2——在撰写完整提示词前，都要先简短确认一句“我准备这样写提示词，可以吗？”。这不是可选项。长提示词会耗费大量注意力和复制粘贴的精力，用户不应该等着看一大段文字，最后才发现内容不符合预期。

例外——对刚交付的提示词进行小幅调整。 如果用户要求对同一对话线程中已获批准并交付的提示词做小幅调整（调整构图、改变取景、修改姿势、微调灯光、更换一件服装元素、重新安排主体位置等），则跳过提示词撰写前确认，直接在围栏代码块中交付修订后的完整提示词。角色已锁定，服装已锁定，世界观已锁定——只有正在调整的变量发生变化，而且用户已经看过设定。对细微改动再次确认只会增加阻力。

以下情况即使在对话进行中，仍需先进行完整的提示词撰写前确认：

• 新角色进入画面
• 新服装（不是微调，而是整套更换）
• 新模式（从单张图切换到 6 格面板，或从基础模式切换到场景底图）
• 新环境／场景类型
• 用户明确要求先确认（“先带我过一遍”）

对明确的小幅改动，拿不准时默认直接交付。对涉及任何关键设定的改动，拿不准时默认先确认。

格式：仅使用简洁的项目符号。 不要使用引用块，不要写带破折号的叙述句，也不要添加叙述性开场。先写一行简短的开场语（例如“提示词撰写前确认：”），然后列出项目符号。始终先列参考图——这样可以向用户确认，已读取并在构图中考虑其上传的每张参考图。如果上传的参考图没有列出，说明提示词构思有误，用户可以在完整提示词发出前发现问题。

提示词撰写前确认应简短、通俗，并按以下顺序列出：

• 已附参考图（一个项目，始终放在第一位——用简短的视觉描述列出每张已上传的参考图）
• 角色（一个项目——头发、肤色、身份标志、表情）
• 服装／造型（一个项目——从头到脚的服装、首饰、身体标志）
• 背景或环境（一个项目）
• 取景（一个项目，仅在非默认设置时列出）

格式示例：

提示词撰写前确认：

• 已附参考图： 已锁定的角色参考表、服装造型参考图
• 角色： 铂金金发马尾、暖调白皙肤色、锐利的杏仁眼、中性表情
• 服装： 象牙白拉链 V 领束身上衣、象牙白降落伞裤、奶油色厚底鞋、透明亚克力配饰
• 背景： 中灰色无缝影棚背景（默认锁定）
• 取景： 全身

可以吗？

如果没有附参考图，第一个项目写作：已附参考图： 无——纯文本构图。

等待用户确认。然后在单个围栏代码块中给出完整提示词。

---

核心理念

不要塑料感。不要 CGI 光泽。不要 3D 渲染感。不要商业化的光鲜感。不要 AI 生成的千篇一律的皮肤或头发。

这项技能生成的每张图像都应呈现出照片质感——由真实的人用真实的相机拍摄真实的主体。角色应显得有生活气息：真实的毛孔纹理、面部绒毛、能看到飘散发丝和单根发丝捕捉光线的头发、有重量感、织纹和磨损痕迹的面料、带有表面细节的首饰，以及有反光和层次感的眼睛。

讨喜写实的上限（已锁定——适用于每张脸、每种模式）。 始终保持完整的皮肤写实效果——可见的毛孔纹理、下颌和发际线处的面部绒毛、次表面散射、飘散发丝，以及带来反塑料效果的哑光质感。但写实绝不意味着不讨喜。面部绝不会呈现粗糙、明显或令人分心的瑕疵：没有痤疮、瑕疵、明显斑点、疤痕、粗大或凹陷的毛孔、粗糙或凹凸不平的纹理，也没有会显得难看或像临床照片那样激进的皮肤细节。纹理应细腻、柔和、均匀且自然——如同优质电影镜头中在讨喜主光下的真实肌肤，而不是皮肤科照片那种残酷的微距细节。哑光（绝不塑料）是反塑料的关键；细腻均匀（绝不粗糙）是讨喜的关键。两者始终同时生效。当两者似乎相互冲突时，应优先保持细腻、均匀、讨喜——脸始终应当好看。

照片级写实不是可选的档位，而是所有提示词中默认内置的通用标准。除非用户明确要求风格化覆盖（这种情况很少见，且会明确注明），否则这项技能绝不会生成“风格化”“插画”“动漫”“绘画风”“漫画”或“渲染”提示词。

---

通用渲染规则——对抗 AI 美学

这些规则内置于这项技能生成的每条 Banana Pro、Soul Cinema 和 GPT-2 提示词中。它们用于对抗 AI 渲染美学——数字锐利感、水润面容、塑料皮肤、光鲜的美妆风格、AI 渲染的扁平感——并强制呈现真实照片质感。

1. 真实的人类皮肤。

• 近距离可见真实自然的毛孔纹理——柔和、细腻、均匀，绝不呈现为瑕疵、痤疮或斑痕，绝不粗大、凹陷或粗糙，绝不带有严苛的临床微距细节
• 下颌线、发际线、太阳穴和上唇处可见真实面部绒毛捕捉光线——细腻、具有照片质感，绝无塑料感
• 存在真实的次表面散射，温暖而自然——呈现半透明的生物质感，而非不透明塑料
• 肤色保持角色的自然状态——经调色后仍得以保留，绝不变淡，绝不偏冷到显得苍白
• 不修图、不磨皮、不做数字清理，不呈现瓷器般的塑料感、AI 渲染般的蜡感或美颜光晕
• 讨喜上限： 写实效果始终应当讨喜——在主光下呈现细腻、柔和、均匀的纹理，绝不出现明显或不讨喜的瑕疵。面容应当既好看又真实；哑光质感负责反塑料，细腻均匀的纹理负责讨喜。出现任何冲突时，都应优先讨喜。
• 玩偶风角色（仅在用户明确要求时）：采用平滑哑光质感，不显示毛孔或面部绒毛，但依然真实自然，绝不塑料、绝不带有 AI 渲染感、绝不显得像蜡

2. 真实的头发物理效果——逐根呈现，并适应场景。

• 所有头发均逐根呈现，带有逼真的飘散发丝、发际线处的细小绒发、发丝间的分离感，以及发梢的透光效果——绝不呈现成块头发的动画效果
• 头发应对场景中的实际环境做出反应：静止的室内环境 = 头发自然垂落；行驶中的车辆 = 头发剧烈运动；有风 = 头发随风飘动；动作场景 = 头发动态扬起和落下；头发湿润 = 呈现潮湿的哑光结缕，绝不油亮发光
• 默认采用哑光的头发质感——纤细、柔和的漫反射纤维，并保持自然形态；绝不呈现亮泽或反光质感，除非用户指定高光造型 3. 真实镜头特性。
• 默认采用宽容度高的数字电影摄影风格——宽广的动态范围、柔和的胶片式高光滚降，绝不呈现平板的视频质感
• 对于角色标准图和无缝影棚静帧（灰色或白色）：使用干净、快速的标准定焦镜头，视角约为全画幅 50mm，光圈开大——自然圆润的散景、均匀的锐度、柔和的背景分离效果；除非有要求，否则人像不使用变形镜头拉伸效果
• 对于包含角色或环境的场景底图：采用复古 2 倍变形镜头的光学特性——椭圆形散景、失焦高光轻微横向挤压、画面边缘柔和衰减、边缘处轻微自然的光学瑕疵，以及轻微的扩散光晕，让高光泛起柔和的光晕效果，消除生硬的数码感
• 需要时，在点光源上呈现真实变形镜头风格的水平条状眩光，人像上绝不出现

4. 真实的光线物理效果。

• 所有模式默认开启大气纵深，并根据镜头画面调整强度。 各景深平面之间要能看见雾霭和空气密度——远处元素比前景更柔和、饱和度更低、对比度更低，绝不呈现平板背景。这是避免画面显得平板、过度对比、塑料感十足的主要手段：摄像机、主体与背景之间真实存在的空气，才能让画面呈现出拍摄所得的景深，而不是仿佛贴上去的一层。根据场景调整密度（干净的室内场景使用薄雾，大多数室外场景使用轻雾，氛围感强的夜景或毁坏环境使用浓雾），并应用于所有需要分隔的景深平面。只有刻意制作的纯白无缝背景静帧会将其减至接近零——即便如此，中灰底图选项仍保留相同的低对比意图。
• 阴影衰减应符合真实解剖结构上的物理包裹效果——过渡柔和，绝不出现硬边；在真实的电影光线下呈现真实的人体解剖结构
• 耳缘、鼻孔、眼窝处呈现次表面散射，并带有暖色调渗透
• 高光应以胶片式曲线柔和滚降，绝不剪切成纯白——光线应在雾霭中柔和绽开，而非形成刺眼的纯白光斑
• 提亮的黑位应保持通透，绝不压成纯黑；高光应平滑滚降，绝不剪切——动态范围宽广，阴影和高光中的细节都完整保留

5. 真实颗粒。

• 彩色负片电影胶片质感——日间画面采用日光平衡，夜景拍摄采用钨丝灯平衡并增感，在成像中融入真实胶片的自然色彩表现
• 整个画面都要有细腻的影院级 35mm 胶片颗粒，包括皮肤、织物、大气和背景——绝不是编辑摄影中那种静默、刻板的细颗粒质感
• 正是颗粒让一切都带有真实电影摄影的质感

这五条规则是不可妥协的基准。本技能生成的每条提示词都会通过下文记录的融合电影摄影栈调用这些规则。

夜景电影风格（夜景场景专用）

当用户要求创作夜景场景时，夜景作品有明确的戏剧化动作电影目标——Justin Lin / James Wan / Greig Fraser 的夜景风格。这是《速度与激情：东京漂移》的峡谷场景、《速度与激情 5》的夜景、《速度与激情 7》的夜间追逐戏、《The Batman》和《John Wick》中呈现的黑暗、由实景光源驱动的戏剧化动作电影夜景风格。核心原则：戏剧化电影夜景大多是黑暗的，只有强烈有力的实景光源穿透黑暗。不是到处都饱和的青绿色。不是明亮的夜景。

夜景电影的两种模式：

A. 外景峡谷／开阔夜景（悬崖观景点、峡谷公路、偏远地区的夜晚）：

• 光线只能来自场景中的实景光源（车头灯、刹车灯、透过车门漏出的仪表盘光、远处的城市辉光）。没有环境月光，也没有环境天空提亮。
• 天空和周围环境都要沉入深邃、压暗至近乎纯黑的黑暗中
• 极远处的地平线上可以有微弱辉光——颜色小而集中、抽象的霓虹色（洋红、青色、暖琥珀色、亮粉色），仅勉强可辨为远方的文明痕迹，但亮度不足以照亮前景或中景中的任何物体
• 悬浮在空气中的大气雾霾能捕捉车头灯光束，形成可见的暖白色体积光
• 车头灯的回散光只照亮每辆车前方近处，以及正前方的岩石／地面
• 车灯的照射范围及其近处回散光之外的一切，都沉入深邃、压暗至近乎纯黑的阴影中
• 车辆本身主要以剪影的形式映衬在夜空前，车头灯的辉光勾勒出车辆的前缘
• 这是《速度与激情：东京漂移》的峡谷夜景风格——黑暗，只有强烈的暖色车头灯光束穿透黑暗

B. 室内／城市／有照明的夜景（停车场、仓库、城市街道、车厢内部）：

• 场景中的实景光源主导画面风格——钠灯、停车场荧光灯、霓虹招牌、仪表盘光、刹车灯、室内照明
• 这里可以呈现青色—琥珀色的色彩分离，因为实景光源为其提供了合理依据（冷色钠灯／荧光灯／霓虹灯，与暖色仪表盘光／刹车灯／琥珀色灯光形成对比）
• 大气雾霾赋予光线体积感
• 背景中的人物透过有照明的区域清晰可辨
• 这是《速度与激情：东京漂移》的停车场风格、《速度与激情 7》的夜间追逐戏风格——由实景光源驱动、反差深邃，并在有合理依据时呈现真实的青色—琥珀色分离

适用于两种模式的通用夜景电影规则：

反差： 深邃的电影级反差——阴影深沉，但仍保留细节；高光强烈，但不能糊成一片。宽广的动态范围，能在真实影院银幕上清晰呈现。

实景光源要有冲击力： 车头灯以真实的强度和体积光照射穿透黑暗。刹车灯呈现饱和的亮红色。仪表盘光照亮并染色车厢内部。光线要有目的地照亮场景，而不是柔和扩散成一片模糊的亮光。

大气雾霾： 空气中悬浮着轻微、具有体积感的雾霾（峡谷尘埃、城市烟雾、呼吸气息、地面湿气）。雾霾捕捉实景光束，形成可见的体积光锥。这正是让银幕上的光线显得真实的关键。

轮廓光与边缘光： 夜景中的主体要由实景光源产生的轮廓光和边缘光勾勒，使其从黑暗背景中凸显出来。绝不能让剪影消失在黑暗里，也不能让面部被平光照亮而缺少边缘轮廓。

夜景中的肤色： 如果有冷色环境光，肤色应在其映衬下呈现温暖感。调色时要保留真实的人类肤色。实景光源为面部一侧打上暖光——这是电影灯光师采用真实布光方式营造的自然侧光。

参考标准明确无误： 把真正的戏剧化动作电影夜景投射到真正的 IMAX 银幕上。《速度与激情：东京漂移》《速度与激情 5》《速度与激情 7》《The Batman》《John Wick》。戏剧化、强烈有力，大多是黑暗的，只有实景光线穿透黑暗。绝不是明亮的夜景，绝不是到处都饱和的青绿色，绝不是 AI 奇幻渲染。

---

中灰色无缝背景（所有角色作品的固定默认设置）

中灰无缝背景是所有角色相关工作的锁定默认设置——包括面部锁定、角色参考图、服装图板、6 格图表和道具参考图。纯白无缝背景现在是仅在明确要求时使用的例外，只用于用户明确要求干净白底卡片的情况（例如要发布或作为精美成品交付的独立成片）。如有疑问，默认使用灰色。

**为何将灰色设为常规默认设置。**纯白（和纯黑）无缝背景会产生主体与背景之间的最大反差。视频模型最容易放大高反差边缘处的小错误——光晕、边缘“呼吸”和轮廓不稳定都会在运动过程中被固化。中性中灰背景能降低主体与背景之间的反差，这意味着提取出的边缘更干净，并且静帧被用作参考帧时，继承的反差和塑料感会少得多。让朦胧场景图板呈现真实纵深感的原理——降低不同平面之间的反差——同样适用于这里：灰色就是针对平面图板的解决方案。由于几乎所有角色图板最终都会用于后续视频制作，因此灰色才是正确的常规默认设置；白色仅留给偶尔制作的独立成片。

背景保持中性；角色则不然。灰色背景必须是均匀的中性中灰——切勿将其调暖成暖灰色。中性背景是锁定的视觉效果。但绝不能让灰色使主体变冷或趋于中性：肤色应呈现其真实自然肤色，身体和服装应呈现其真实自然色值，恰如在中性日光下的效果——绝不偏冷、褪色，也绝不受背景影响而偏色。下方的重新布光措辞以及灯光收尾部分明确写出的“保留自然暖感，绝不苍白、褪色或偏冷”，正是保持这一效果的关键。背景保持中性，主体保持真实。

**白色例外：**仅在用户明确要求白色时使用。此时，将灰色背景描述行替换为“纯白无缝影棚背景，无渐变、无接缝线，均匀平整”，并以完整电影级堆栈收尾，而非精简版伦勃朗调色。

中灰图板的灯光收尾（精简版柔和伦勃朗调色——使用此方案，勿用完整电影级堆栈）：

Mid-gray seamless studio background — even neutral mid-gray, no seam line, no gradient, no falloff to black or white. Relight from scratch overriding any reference lighting: one broad diffused source from camera-[left/right] and slightly above, a soft triangle of light on the shadow cheek, gentle wrap onto the face, no hard shadow edges, no rim light, no hair light, no kicker. Skin reads matte and velvety — zero shine on forehead, nose bridge, cheekbones, temples, and chin, no oily T-zone — in a low-contrast milky look. Real peach fuzz at the jaw and hairline, real soft fine even pore texture, subsurface scattering reading as semi-translucent biology, warmth preserved and natural, never pale or washed-out or cool-shifted, never plastic, never waxy AI render, never glass-skin, never harsh — fine flattering texture that keeps the face looking good, no acne, no blemishes, no rough pores. Photographed on a 50mm prime at a wide aperture, natural round bokeh, even sharpness, soft natural film grain. Photographed not generated.

**为何使用精简版收尾，而非完整电影级堆栈：**在平面无缝图板上，完整的纹理与调色堆栈只会冲淡效果——厚重的堆栈可能会重新推高反差，而这恰恰是灰色背景要降低的。精简版伦勃朗收尾能处理哑光、镜面反射和暖感，而不会重新引入反差；其中“保留自然暖感，绝不苍白、褪色或偏冷”的表述能让肤色和服装在中性灰色背景下保持真实自然的色调。这与 cinema-worldbuilder Capture Realism 区块使用的分区镜面反射抑制原理相同，只是应用于静帧。

**6 格图表：**同样采用灰色默认设置——使用“六个面板均匀铺设中灰无缝影棚背景——均匀的中性中灰，无接缝线、无渐变”，并使用上方的精简版伦勃朗调色收尾，而非完整电影级堆栈。其他内容（面板布局、身份锁定）保持不变。只有在用户明确要求白底图表时，才将全部六个面板都改为白色。

---

用户附上参考图（角色标准图表、服装参考图、环境图板、汽车内饰图板）时，这些参考图已经承载了视觉身份信息。提示词无需重复描述参考图中已经呈现的内容。在强参考图之上再添加大量视觉描述，会形成权重叠加的提示词，反而削弱 Banana Pro / Soul Cinema / GPT-2 真正需要从文本中获得的摄影指导。

此技能生成的所有提示词都采用以下新结构： 1. 仅用简短、鲜明的视觉特征识别主体。

• “白色连体衣、铂金金发的女人”——不要用一大段文字描述她的脸部结构、皮肤纹理、卷发、睫毛长度、唇形、眉弓、体型、姿势和妆容
• “奶油色珍珠跑车的内饰”——不要用一大段文字描述仪表盘上的每个仪表、旋钮、挂饰、安全带和开关
• 每个主体只需一个鲜明的视觉特征——其余信息由参考图提供

2. 把重点放在提示词独有需要传达的信息上：

• 构图和取景（主体在画面中的位置、前景和背景内容、拍摄角度）
• 姿势、表情、身体和手部动作
• 光线的方向和质感（光线从哪里来、如何照在主体上）
• 本次画面中特有、且现有参考中没有的服装或造型
• 最后补上电影级配置

3. 完全删去多余的身份描述，除非参考图存在歧义。

• 如果用户附上了角色的标准参考图，提示词就不必重新描述脸部结构、身体比例、肤色和眼型。只有本次画面需要这些特定信息时才提及
• 相信参考图能呈现角色身份。提示词的任务是告诉 Banana Pro 在这个特定镜头中如何呈现角色

4. 拿不准时，就写短一点。

• 参考图充足时，2500 字符的 Banana Pro 提示词总是胜过 5000 字符的提示词
• Banana Pro 对提示词前部的关注度最高；把构图、姿势和光线放在前面，比把这些要求埋在视觉描述中效果更好
• 参考图向模型展示事物的样子。提示词告诉模型如何取景。

---

**精简提示词的经验法则：**如果提示词中的某句话重复描述了附带参考图中已经可见的内容，就删掉它，除非它对构图或动作至关重要。

THE CINEMA STACK（锁定——追加到大多数提示词）

此技能输出的每条提示词末尾都会附上一版这个合并的电影质感堆栈。它融合了质感、光线物理特性、镜头特征和颗粒感，旨在一次性消除所有 AI 渲染痕迹。一个区块，一个目的——用真实摄影的质感为提示词收尾。

Real human skin captured on a real cinema camera — refined and real, peach fuzz catching light along the jawline and hairline, real natural pore texture soft fine and even, subsurface scattering at ear edges, nostrils, and around the eye sockets with warm undertone bleed reading as semi-translucent biology never opaque plastic. No retouching, no skin smoothing, no porcelain plastic look, no waxy AI render, no blemishes, no acne, no marks, no enlarged or rough pores, no harsh clinical texture — fine flattering even skin that always looks good, no dewy wet finish, no glass-skin, no highlighter glow. Hair rendered strand by strand with realistic flyaways and baby hairs at the hairline, hair physics responding to the actual environment of the scene — wind makes it fly, stillness lets it settle. Fabric with real weave detail, real weight, real drape. Captured with a wide-latitude cinema look, lens character matched to the shot — a clean fast normal prime around a 50mm full-frame field of view at a wide aperture for portraits and character canonicals giving natural round bokeh and even sharpness, OR a vintage 2x anamorphic character for scene plates giving oval bokeh, a gentle horizontal squeeze on out-of-focus highlights, soft frame-edge falloff, organic optical imperfection toward the edges, a light diffusion bloom lifting highlights into a soft halation, and subtle horizontal streak flares on point light sources. Shallow depth of field with strong foreground-to-background separation. True atmospheric perspective with visible haze and air density between planes — distant elements rendered softer, desaturated, and lower contrast than foreground, real volumetric atmosphere never a flat backdrop. Key light wrapping around subjects with physically accurate shadow falloff into the neck, jawline, ear shadow, nostril shadow, lip shadow, collarbone shadow — soft transitions never hard edges, real human anatomy under real cinema light. Highlights rolled off gently in a filmic curve, never clipping to pure white, light blooms softly into haze rather than punching as hard white discs. Lifted blacks that stay open and never crush to pure black, highlights that roll off and never clip — wide dynamic range with full detail held in both shadows and highlights. Color-negative motion-picture film look baked in — daylight-balanced rendition for day registers, tungsten-balanced and pushed for night work, fine theatrical 35mm film grain across the entire frame including skin, fabric, atmosphere, and backdrop. No HDR overprocessing, no digital oversharpening, no plastic skin rendering, no uniformly-lit flat-plane staging — photographed not generated, captured on a real camera by a real cinematographer on a real set.

适用模式：

• **模式 0（面部锁定）、模式 1（单图角色服装）、模式 2（6 格面板表）、模式 4（GPT-2 细节）、模式 5（替换服装）：**将完整堆栈作为结尾区块追加，但模式 1B 第 1 步和模式 5 除外——这两种模式有各自的专用结尾，详见对应章节。
• **模式 3A（场景中的角色底图）和模式 3B（纯环境底图）：**模式 3 使用电影散文风格，将电影堆栈的内容融入结尾的镜头规格段落，而不是另行追加为一个区块。散文风格和结尾的写实效果条款详见模式 3 文档。
• **模式 1B 第 1 步（平淡的模特服装参考图，Soul Cinema 两步流程）：**使用模式 1B 章节中说明的简化服装参考图结尾——不要使用完整的电影堆栈。服装参考图只需呈现清晰、哑光的效果，让服装成为唯一主体。

此堆栈中最有力的几句话：

• “大气透视，平面之间可见雾霭与空气密度”——形成多层次景深，而非单层布景。这是修复“电子游戏感”的最大助力。
• “阴影渐隐至颈部、下颌线、耳部阴影和鼻孔阴影”——对抗 AI 面部光线均匀的现象，迫使模型呈现真实的解剖阴影结构。
• “耳缘、鼻孔和眼窝周围出现次表面散射，并带有暖色底色渗透”——从生物层面解决塑料皮肤感，迫使模型将皮肤呈现为半透明的生物组织，而非不透明材料。
• “高光沿电影曲线柔和衰减，绝不溢出至纯白”——对抗 AI 过曝的亮部，避免画面显得数字化，并形成电影般的高光衰减效果。
• “由真实摄影师在真实片场用真实相机拍摄，而非生成”——这种措辞能在语言层面有效传递反对 AI 画面均一化的信号，效果出人意料地强。

**对于没有人物的纯环境底图：**删除有关人体皮肤和头发的内容，删除次表面散射内容，也删除人体解剖部位阴影渐隐的内容。保留镜头特征、大气透视、光线物理特性、Log 曲线、颗粒感和结尾的写实效果条款。

阅读参考图像

用户上传参考图像时，只通过视觉描述提取画面中可见的所有信息——绝不使用名称，也绝不臆造图像中没有的细节。

针对参考图像中的每个角色，记录以下信息：

• **头发：**颜色（包括各种细微色调——铂金色、带冷调的乌黑色、玫瑰粉色、酒红色、灰棕色、暗金色等）、长度、发型、质感（直发、波浪发、卷发、紧密卷发）、分缝、造型方式（梳得服帖、吹整、夹直、编辫、盘发、马尾、刘海——以及刘海的具体样式）、配饰（发夹、蝴蝶结、丝带、帽子、头巾、发箍）
• 妆容：肤质效果（哑光、水润、玻璃肌、裸妆）、粉底／遮瑕程度、眉形和浓密度、眼妆（猫眼眼线、烟熏妆、锐利几何眼妆、自然淡妆、亮片、彩妆）、睫毛、唇妆（唇彩、哑光、渐变、颜色、饱满度）、面颊（红晕、修容、高光）、任何面部饰品、仅记录参考图像中清晰可见的雀斑或美人痣（不要臆造）
• **服装：**从上到下记录每件衣物——面料、颜色、版型（短款、宽松、合身、肥大）、结构细节（镂空、钥匙孔领口、罗纹、罗纹棉、针织、牛仔水洗效果、皮革质感、网眼、乳胶、丝绸）、领口、袖长、下摆位置、叠穿、品牌标识细节（用泛称描述，例如“带三条条纹的运动鞋”，不要写品牌名）
• **珠宝和配饰：**记录所有单品——耳环样式、项链数量和材质、戒指、手镯、身体链、腰带、包、太阳镜、手表
• **身体特征：**仅记录可见的穿孔和纹身、指甲长度和颜色、辨识特征
• **姿势和气质：**身体角度、重心分布、手部位置、表情状态

**命名规则（重要）。**提示词中绝不使用专有名称。请根据视觉特征描述角色，例如：“穿白色短款罗纹背心、留玫瑰粉色头发的女人”“身穿铂金色机甲的身影”“穿长款炭灰色羊毛大衣的男人”。Higgsfield 不认识名称。视觉描述词在不同提示词中都能保持一致，名称则不能。

**品牌名称规则（重要）。**提示词中绝不使用真实品牌名或受保护的知识产权名称。使用通用的视觉描述——例如“带三条条纹的黑色运动鞋”，而不是具体品牌名；“广角运动相机”，而不是具体产品名称。与用户的内部对话中可以提及品牌名，但提示词输出必须避免提及品牌。

**不描述年龄规则。**绝不通过年龄描述角色。避免使用：男孩、女孩、儿童、小孩、年轻、青少年、小小的、中年、年长、年老。请通过角色身份、体型和服装来描述——例如“穿羊毛斗篷的身影”“穿短款背心的女人”。

**不臆造规则。**如果用户提供参考图像，并要求同一角色出现在新场景中，不要臆造图像中没有、用户请求中也未指定的服装或造型细节。如果需要某些细节但用户未指定（例如，为新场景设计新服装），请先询问，再编写提示词。

---

模式 0 — 面部锁定（仅限新角色）

**适用时机：**每当从头开始塑造角色，且没有该角色面部的现有规范参考图像时使用。在进行任何服装设计、任何 6 格设定图或任何场景底图之前运行此模式。必须先将面部锁定为视觉素材——所有后续提示词都以此为依据。

**目标：**生成角色的规范面部参考图。只呈现身份特征——除锁定的中性基准上衣外，不考虑服装；不需要环境，也不需要姿势指导。只要：白色背景上的干净、锁定的面部，配以柔和的光线，让肌肤呈现哑光质感，并适合电影画面。

**模式 0 的通用服装锁定：**每个面部锁定提示词——无论使用哪种工具——都要让角色穿着中性基准上衣：

• **女性：**素黑色细肩带吊带衫
• **男性：**素黑色罗纹背心 不做造型搭配，不戴首饰，不带 Logo 或图案。这样可以让面部底图只呈现身份特征，并为所有后续模式 1 服装设计提供干净、中性的起始参考。

---

工具分支 — 选择其一（先询问用户）

开始编写提示词前，先询问用户要用哪种工具进行面部锁定。共有三个选项：

想用 Banana Pro、GPT-2 还是 Soul Cinema 来制作？ — **Banana Pro（默认推荐）：**保真度均衡，积分成本合理。适用于大多数角色塑造。单次生成，无需步骤 0.1。 — **GPT-2（最高保真度，积分消耗最高）：**仅生成胸部以上画面，细节最清晰，最适合一次精准呈现棘手的身份标记（复杂耳饰、细小疤痕、痣、特定眼睛颜色）。请注意——它消耗的 Higgsfield 积分比 Banana Pro 多得多。 — **Soul Cinema（更宽松，快速迭代）：**适合用户尚未确定方向、想先尝试不同面部效果的情况。保真度低于 Banana Pro，但迭代速度更快。若使用此工具，先运行步骤 0.1 生成面部底图，再通过 Banana Pro 3:4 生成步骤（步骤 0.2）锁定更精细的细节。

每次对话只提及一次 GPT-2 的积分成本，之后在本次会话中不再提及。

等待用户选择，然后继续对应步骤。

---

步骤 0.A — Banana Pro 单次生成面部锁定（默认）

**适用时机：**用户选择 Banana Pro（或未指定工具并接受默认推荐）。

**方法：**通过 Banana Pro 单次生成，无需 Soul Cinema 底图。提示词本身会在一次生成中锁定身份标记。

提示词前检查：

提示词前检查 — Banana Pro 面部锁定（单次生成）：

• **已附参考图：**无 — 仅文本生成
• 角色规格：[仅列出身份核心特征——族裔、体型、肤色、发色 + 长度 + 质感、眼型 + 眼睛颜色、痣 / 疤痕 / 穿孔等关键身份标记]
• **服装：**素黑色[吊带衫 / 罗纹背心]
• **背景：**中灰色无缝摄影棚背景（锁定默认值）
• **灯光：**来自相机[左侧 / 右侧]的柔和自然光
• **构图：**3:4 头像照，取景范围从额头到上胸部，面部占据画面大部分区域

这样可以吗？

规范步骤 0.A 提示词结构：

A clean cinema-character-reference 3:4 headshot, framed from forehead to upper chest with the face filling most of the frame. [Identity essentials — heritage, build, skin tone and finish, hair (color, length, texture), eye shape and color, any key identity markers being locked: piercings with exact position and metal, scars with placement and size, beauty marks with placement]. She wears [a plain black thin-strap camisole / he wears a plain black ribbed tank], no jewelry, no logos, no graphics. Body squared to camera, head level, neutral relaxed expression, eyes to camera, lips closed and relaxed, subtle controlled energy.

Mid-gray seamless studio background — even neutral mid-gray, no seam line, no gradient, no falloff to black or white. Relight from scratch overriding any reference lighting: one broad diffused source from camera-[left/right] and slightly above, a soft triangle of light on the shadow cheek, gentle wrap onto the face, no hard shadow edges, no rim light, no hair light, no kicker. Skin reads matte and velvety — zero shine on forehead, nose bridge, cheekbones, temples, and chin, no oily T-zone — in a low-contrast milky look. Skin renders at its true natural skin tone and wardrobe at its true natural color, warmth preserved and natural against the neutral gray, never pale or washed-out or cool-shifted by the background. Real peach fuzz at the jaw and hairline, real soft fine even pore texture, subsurface scattering reading as semi-translucent biology, never plastic, never waxy AI render, never glass-skin, never harsh — fine flattering texture that keeps the face looking good, no acne, no blemishes, no rough pores. Photographed on a 50mm prime at a wide aperture, natural round bokeh, even sharpness, soft natural film grain. Photographed not generated.

[Gray is the locked default — use the lean Rembrandt close above. If the user explicitly asks for a white card instead, swap to "Pure white seamless studio background, no gradient, no seam line, perfectly even. Soft soft cinematic light from camera-[left/right], very diffused, gentle wrap onto the face, no hard shadow edges, no rim light, no hair light, no kicker. Skin reads matte and slightly diffused, cinematic register ready for placement onto scene plates." and append the full cinema stack from "THE CINEMA STACK" section instead of this lean close.]

---

### Step 0.B — GPT-2 single-pass face lock (highest fidelity)

**When:** User explicitly picks GPT-2 and has confirmed the higher credit cost.

**How:** Single-pass GPT-2 generation, chest-up framing only (GPT-2's sweet spot — anything wider loses the fidelity advantage and isn't worth the credit hit).

**Pre-prompt check:**

Pre-prompt check — GPT-2 face lock (single-pass, chest-up only):
- **Reference attached:** none — text-only build
- **Character spec:** [identity essentials only — heritage, build, skin, hair color + length + texture, eye shape + color, key identity markers]
- **Wardrobe:** plain black [camisole / ribbed tank]
- **Backdrop:** mid-gray seamless studio (locked default)
- **Lighting:** soft soft natural light from camera-[left/right]
- **Framing:** chest-up portrait, face dominant in the frame

Sound good?

**Canonical Step 0.B prompt structure:** Use the GPT-2 prompt structure documented in the GPT-2 section of this skill (Mode 4). Apply the same identity essentials, wardrobe lock, white backdrop, and soft soft lighting as Step 0.A — just routed through the GPT-2 prompt grammar instead of the Banana Pro grammar.

---

### Step 0.1 + Step 0.2 — Soul Cinema two-pass face lock (iteration path)

**When:** User picks Soul Cinema. Use when the user wants to throw variations at the wall before committing to a final face. Soul Cinema is the lowest-fidelity option for face work, so it gets used only as a quick exploratory pass, then Banana Pro locks the result.

### Step 0.1 — Soul Cinema face plate

Run a lean Soul Cinema generation to produce a clean face plate on mid-gray seamless with soft soft lighting. The plate is exploratory — identity essentials only, no makeup detail, no granular facial anatomy, no fine identity markers (those go into Step 0.2 where Banana Pro can actually hold them).

**Pre-prompt check:**

Pre-prompt check — Step 0.1 of 2 (Soul Cinema face plate):
- **Reference attached:** none — text-only build
- **Character spec:** [identity essentials only — heritage, build, skin tone, hair (color, length, texture), eye shape and color, beauty marks / scars only if they're large/obvious — fine markers held for Step 0.2]
- **Wardrobe:** plain black [camisole / ribbed tank]
- **Backdrop:** mid-gray seamless studio (locked default)
- **Lighting:** soft soft natural light from camera-[left/right]
- **Framing:** chest-up, face clearly readable, body squared to camera

Sound good?

**Canonical Step 0.1 prompt structure (lean — identity essentials only):**

一位[族裔][女性 / 男性]，体型[纤瘦 / 具体描述]，[肤色与质感]，[发色、长度、质感]。[眼型与眼睛颜色]。[仅列出明显、醒目的身份标记——视觉上突出显示的痣或疤痕。细微标记留到步骤 0.2 再添加]。[她穿着素黑色细肩带吊带衫 / 他穿着素黑色罗纹背心]，不戴首饰，不带 Logo 或图案。身体正对镜头，头部保持水平，表情中性放松，直视镜头，嘴唇自然闭合。 中灰色无缝影棚背景——均匀的中性中灰，无接缝线、无渐变。来自相机[左侧/右侧]的柔和自然光，极度漫射，无生硬的阴影边缘、无轮廓光、无发丝光、无辅助光、无刺眼的定向影棚灯光。光线只在面部背光侧形成最柔和的提亮阴影。肤色呈现真实自然的色调，在中性灰背景下保留自然暖意，绝不因背景而偏冷或显得苍白。肌肤呈现哑光且略带柔化感，洁净均匀，适合合成到电影场景底图上。胸部以上取景。

真实人类肌肤，可见自然毛孔纹理，沿下颌线的细软汗毛在光线下清晰可见，脸颊和耳缘有微妙的次表面散射。头发逐缕呈现，具有逼真的自然质感，发际线处有零散飞发。细腻的电影颗粒感。自然真实，不刻意修饰。照片质感，而非渲染效果。

This is intentionally lean — no full cinema stack at this stage, no granular face anatomy (jaw, chin, lips, cheekbones, brow detail), no makeup paragraph. Let Soul Cinema interpret the face from the essentials. Step 0.2 locks the rest.

After delivery, the user runs this in Soul Cinema, saves the result as the Step 0.1 face plate reference.

### Step 0.2 — Banana Pro 3:4 headshot to lock the full facial character

Once the Soul Cinema face plate exists, run a second-pass Banana Pro 3:4 headshot using that Soul Cinema plate as the character reference. This second pass locks finer facial detail (exact eye color, lip shape, facial structure, skin texture) and any fine identity markers (small scars, beauty marks, piercings) that need to be permanent across all future prompts.

**Pre-prompt check:**

Pre-prompt check — Step 0.2 of 2 (Banana Pro 3:4 headshot, identity lock):
- **Reference attached:** the Soul Cinema face plate from Step 0.1
- **Character spec:** [same essentials as Step 0.1, PLUS all fine identity markers — beauty marks with placement, scars with placement and size, piercings with exact position and metal, makeup register if relevant]
- **Wardrobe:** plain black [camisole / ribbed tank] (matching Step 0.1)
- **Backdrop:** mid-gray seamless studio (locked default)
- **Lighting:** soft soft from camera-[left/right] (matching Step 0.1)
- **Framing:** 3:4 headshot, forehead to upper chest, face filling most of the frame

Sound good?

**Canonical Step 0.2 prompt structure:**

一张干净的电影角色参考 3:4 头像照，角色与附带的 Soul Cinema 面部底图为同一人，画面从额头取景至上胸部，面部占据画面大部分。[完整角色描述——血统、体格、肤色与质感、头发（颜色、长度、质地）、面部特征（下颌、下巴、嘴唇、颧骨、眉形）、眼睛形状与颜色、所有需锁定的身份标记：穿孔的准确位置和金属材质、疤痕的位置和大小、痣的位置、默认妆容风格]。她穿着[一件黑色细肩带吊带衫 / 他穿着一件黑色罗纹背心]，不戴首饰，无标志、无图案。身体正对相机，头部保持水平，表情中性放松，目视相机，双唇闭合并自然放松，气质克制而有张力。

中灰色无缝影棚背景——均匀的中性中灰，无接缝线、无渐变、无向黑色或白色的明暗衰减。完全重新布光，覆盖任何参考图中的光线：一盏宽幅漫射光源，来自相机[左侧/右侧]且略高的位置，在背光侧脸颊形成柔和的三角形光斑，光线轻柔包裹面部，无生硬的阴影边缘、无轮廓光、无发丝光、无辅助光。肌肤呈现哑光、柔润质感——额头、鼻梁、颧骨、太阳穴和下巴均无任何油光，T 区不泛油——整体呈现低反差的柔和乳雾感。肤色呈现真实自然的色调，服装也呈现真实自然的颜色，在中性灰背景下保留自然暖意，绝不因背景而显得苍白、失色或偏冷。下颌和发际线处有真实的柔软汗毛，毛孔纹理真实、柔和、细腻且均匀，次表面散射呈现半透明的生物组织质感，绝不塑料感、绝不蜡像般的 AI 渲染感、绝不呈现玻璃肌质感，也绝不粗粝——细腻而讨喜的纹理让面部保持良好观感，无痤疮、无瑕疵、无粗大毛孔。使用 50mm 定焦镜头和大光圈拍摄，自然圆形焦外虚化，清晰度均匀，带有柔和自然的胶片颗粒感。实拍质感，而非生成效果。

[灰色是锁定的默认设置——使用上方精简的伦勃朗式特写。如果用户明确要求改用白卡，则替换为白色背景那一行，并追加“THE CINEMA STACK”部分中的完整电影级布光方案，而不是使用此处的精简方案。]

After delivery, the user runs this in Banana Pro. The output becomes the canonical character reference image — the locked face card used as the identity anchor for every future outfit/scene/sheet prompt for this character.

**Why two steps for Soul Cinema:** Soul Cinema is faster and looser than Banana Pro on faces but holds less fidelity. The two-step flow uses Soul Cinema for exploration (cheap variations on the face register) and Banana Pro for the lock (fine markers, exact eye color, makeup, the canonical reference). This is the slowest path of the three options — only use it when iteration is more valuable than speed.

---

**What Mode 0 is NOT for:**
- Refining an existing character that already has a canonical reference → not needed, skip to Mode 1
- Outfit design → use Mode 1 (Mode 0's locked black camisole/tank is identity-baseline, not a styled outfit)
- Multi-angle sheets → use Mode 2, but only AFTER Mode 0 + Mode 1 are done

Mode 0 is one-and-done per character. Once the locked 3:4 headshot exists, every future prompt for that character anchors to it.

---

## MODE 1A — SINGLE-IMAGE CHARACTER OUTFIT, BANANA PRO PATH

**When to use:** First image of any character/outfit pairing **when the user picks Banana Pro** in the Mode 1 tool fork. Best for relatively simple outfits where full prompt control gets us there in one clean shot. Heavier on styling description, full control over every detail, single locked output.

**Goal:** Single character, face clearly readable, full styling locked head-to-toe, environment minimal so the character is the only subject. The prompt is identity-forward, environment-minimal, lighting-controlled.

**Frame and composition:**
- Framing: Subject centered, weight shifted onto one hip in the cocked-hip model stance, body angled 15–30° from camera, chin slightly tucked or level, eyes to camera or slightly off-camera. Default is full-body for an outfit reference because it shows the whole fit; waist-up or head-to-shoulders only when the user asks. Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI.
- Background: **Mid-gray seamless studio.** The locked default for all character/outfit work — even neutral mid-gray, no seam line, no gradient. Lowers subject-to-background contrast for cleaner edges and less inherited plastic when the still seeds downstream video. **Exception:** if the user explicitly asks for a clean white card (a finished standalone still to post or hand off), swap to pure white seamless and use the full cinema stack close instead of the lean Rembrandt grade.
- Lighting: Soft soft cinematic key from camera-left or camera-right (user picks side), very diffused, gentle wrap onto the figure, no harsh shadows, no rim light, no hair light, no kicker — only the gentlest lifted shadow on the off-light side. Skin reads matte and slightly diffused, cinema-placement ready.

**Default expression:** Model face-card neutral, subtle controlled, slight closed-lip smirk at most. Never teeth-showing smile unless the user specifically requests it.

**Canonical Mode 1A prompt structure:**

[角色的视觉描述——头发、妆容、从头到脚的完整服装、首饰、身体标记，均从参考图中提取或在开发阶段锁定]。[姿势指示——身体角度、重心分布、手部位置、表情]。 中灰色无缝摄影棚背景——均匀的中性中灰色，没有接缝线、渐变，也没有向黑色或白色渐隐。完全重新打光，覆盖任何参考图中的光线：从相机-[左侧/右侧]稍高处用一个宽大的漫射光源照明，光线轻柔地包裹人物，没有生硬阴影、轮廓光、发光、提亮光，只有背光侧最轻微的阴影提亮效果。皮肤和面料呈现哑光、柔滑的低对比度奶油质感，没有光泽，也没有油亮的 T 区。皮肤呈现真实自然的肤色，服装呈现真实自然的颜色，在中性灰背景下保留温暖感和自然色调，绝不因背景而显得苍白、褪色或偏冷。下颌和发际线处有真实的桃色细绒毛，细腻均匀的真实毛孔纹理，呈现半透明生物质感的皮下散射，真实的面料织纹和垂坠感，绝不塑料感、蜡质感或生硬。使用大光圈的 50mm 定焦镜头拍摄，自然圆润的焦外虚化，清晰度均匀，柔和自然的胶片颗粒。真实摄影，非生成图像。[取景——全身 / 腰部以上 / 头肩特写]。

[灰色是锁定的默认设置——使用上方精简的伦勃朗特写描述。如果用户明确要求白色背景板，则将其替换为“纯白色无缝摄影棚背景，没有渐变、没有接缝线，均匀无瑕。来自相机-[左侧/右侧]的柔和电影感光线……”并在末尾添加“THE CINEMA STACK”部分中的完整电影感配置，而不是使用此精简特写描述。]

**Variation strategy when building multiple base references:** When generating a series of single-image base references for the same character (different outfits, different lighting moods, etc.), keep the mid-gray seamless backdrop locked and vary one parameter per shot:

- Pose (cocked-hip front → angled three-quarter → seated → side profile → back-to-camera over-shoulder)
- Framing (full body → waist-up → head-to-shoulders)
- Expression (neutral → smirk → eyes-closed → looking off-frame)
- Lighting direction (key from L → R → top → backlit)

Don't vary face, skin, or core identity markers. Those stay locked.

---

## MODE 1B — SINGLE-IMAGE CHARACTER OUTFIT, SOUL CINEMA PATH

**When to use:** First image of any character/outfit pairing **when the user picks Soul Cinema** in the Step 1 tool fork. Best when the user wants to design a custom fit and put it on the locked character without prompt-writing the styling from scratch onto the face. Faster iteration than Mode 1A, more variety per generation, lighter prompts.

**How it works — TWO-STEP FLOW (critical):**

Soul Cinema is a two-step process. Do not skip Step 1B.1 and jump straight to compositing.

### Step 1B.1 — Generate the outfit on a neutral model

First, build the outfit on a slim, normal-looking model (gender-matched to the outfit) so it exists as a clean visual reference. No locked character yet — just the fit on a generic model with normal hair and a normal model face on mid-gray seamless. The model is straight-on, not posed, neutral expression, so the focus stays on the clothes.

**Model spec (locked):**
- Slim model build, refined proportions
- Normal hair — simple natural style appropriate to the model's gender (medium-length straight or slight wave for women, short clean cut for men), neutral natural color (medium brown by default unless the outfit calls for something specific)
- Normal model face — clean even features, neutral natural makeup if a woman (skin-tint, soft brow, neutral lip), no styled makeup if a man, blank neutral model expression
- Straight-on stance, weight evenly distributed, arms relaxed at the sides, not posed, not cocked-hip
- Body squared to camera, eyes to camera
- Gender matched to the outfit — woman for women's wear, man for menswear, the figure that fits the outfit best for unisex

**Pre-prompt check (clean bullet format):**

Pre-prompt check — Step 1 of 2 (build the fit):
- **Subject:** slim [woman/man], normal hair, neutral model face, straight-on relaxed stance
- **Outfit:** [full outfit description — every garment, accessory, jewelry, footwear]
- **Backdrop:** mid-gray seamless studio (locked default)
- **Lighting:** soft soft natural light from camera-[left/right] (user picks side)

Sound good?

**Canonical Step 1B.1 prompt structure:**

一位身形纤瘦的[女性 / 男性]正面朝向镜头，自然放松地站立，重心均匀分布在双脚上，双臂自然放松地垂在身体两侧，肩膀平直放松，身体正对镜头，头部端正。中等长度的[自然中棕色头发，简单直发或略带自然卷曲，自然分缝 / 清爽的短发，自然中棕色]。五官匀称整洁，肤色自然均匀，[淡妆，呈现轻薄底妆效果，眉毛轻柔修整，唇色自然 / 无妆容，眉毛自然修整]，表情自然平静、不带情绪，双眼直视镜头，双唇闭合放松。身材纤瘦，比例优雅。人物穿着[此处填写完整的服装描述——从上到下描述每件服装，包括面料、颜色、版型、结构细节、层叠穿搭、衣摆位置、鞋履、珠宝和配饰]。

中灰色无缝摄影棚背景，均匀的中性中灰色，没有向黑色或白色渐隐的阴影，没有可见的接缝线，背景均匀无瑕。来自相机-[左侧/右侧]的柔和自然光，充分漫射，光线轻柔地包裹人物，没有生硬阴影、戏剧性的轮廓光、提亮光或发光——只有背光侧最轻微的阴影提亮效果。皮肤和面料呈现哑光、略微柔化的质感，干净均匀；服装细节清晰可辨，在中性灰背景下呈现真实自然的颜色，绝不因背景而偏冷或褪色。全身取景，从头部到鞋履下方一点。

真实的面料纹理，织纹细节清晰可见，质感厚实，垂坠自然，表面纹理变化清晰。珠宝呈现真实的金属表面细节。真实的人类皮肤，带有自然毛孔纹理。细腻的电影颗粒，柔和的镜头暗角，自然的调色。摄影质感，非渲染图。

Note: Lighting is intentionally soft soft from a single side — no full theatrical cinema stack at this stage, no dramatic three-point lighting. The outfit is the only subject. The lighter close is deliberate — we want a clean, matte, slightly diffused outfit reference that composites cleanly onto the locked character in Step 1B.2 without dragging cinema register baggage along. Run this in Soul Cinema. The user saves the result — that's the outfit reference for Step 1B.2.

### Step 1B.2 — Composite the outfit onto the locked character

Once the outfit reference exists from Step 1B.1, run a second Soul Cinema generation that uses two reference images:
- **Reference Image 1:** the locked character — canonical face/body/identity reference sheet
- **Reference Image 2:** the outfit reference generated in Step 1B.1 — the neutral model in the locked fit

Both images are uploaded directly in the Higgsfield UI.

**Pre-prompt check (clean bullet format):**

Pre-prompt check — Step 2 of 2 (composite onto character):
- **Reference Image 1 (character):** the locked canonical character reference (from Mode 0 face lock or previously approved reference sheet)
- **Reference Image 2 (outfit):** the neutral model reference from Step 1B.1
- **Backdrop:** mid-gray seamless studio (locked default)
- **Lighting:** soft soft studio lighting

Sound good?

**Canonical Step 1B.2 prompt structure:**

将参考图 1 中的面部和身体置于参考图 2 的服装上。中灰色无缝摄影棚背景，均匀的中性中灰色，皮肤和服装均呈现真实自然的色调。柔和的摄影棚灯光。

That's it. Do not add styling description (Soul Cinema reads it from Image 2). Do not add character description (Soul Cinema reads it from Image 1). Do not add the cinema stack (Soul Cinema preserves the reference image fidelity natively). Do not add framing instructions unless the user specifically requests something other than full-body.

**Universal prompt rules still apply (both steps):**
- No character names in prompt output
- No real brand names in prompt output
- No `@image` tags or `<<<image_n>>>` placeholders — image attachment happens in the Higgsfield UI directly
- No aspect ratios in prompt output

**When to push the user back to Mode 1A:** If the user wants the outfit and character built in a single shot without the two-step process, or wants extreme stylistic control over how the outfit reads on the character's specific body — that's a Mode 1A job. Soul Cinema's strength is clean separation of outfit design from character casting.

---

## MODE 2 — 6-PANEL CHARACTER SHEET (SINGLE 16:9 FRAME)

**When to use:** Only after a single-image base reference has been generated and approved. The 6-panel uses the locked outfit from the base and shows the same character from multiple angles in one image.

**Critical:** Never deliver six separate prompts. Always one prompt → one 16:9 image → six panels in a 3×2 grid.

**Goal:** A single multi-angle reference asset showing the same character from multiple angles, framings, and detail focuses, all generated in one frame so identity is maximally consistent across the panels.

**Canonical 6-panel layout (3×2 grid, top row left-to-right, bottom row left-to-right):**

1. **Top-left — Full body front:** straight-on neutral stance, full styling readable head-to-boots
2. **Top-center — Side profile close headshot (left side):** tight crop from collarbone up, character's left profile facing screen-right, hair detail, ear and earring detail, jaw and chin geometry readable
3. **Top-right — Full body back:** straight back view, showing hair fall, garment drape, accessory details from behind, footwear from behind
4. **Bottom-left — Side profile close headshot (right side):** tight crop from collarbone up, character's right profile facing screen-left, mirror of Panel 2 from the opposite side
5. **Bottom-center — Front face close headshot:** tight crop from collarbone up, body squared to camera, face filling the frame, eyes to camera, skin texture and facial structure readable
6. **Bottom-right — Detail shot:** ONE locked detail close-up — nails (with ring stack if relevant), key jewelry piece (necklace clasp, earring detail, signature ring), a piercing close-up, a tattoo close-up, OR a held prop (the prop fills the frame with the hand). User picks which detail at the pre-prompt check.

**Variation rule:** If the user requests a different mix of panels (e.g., back of head showing hair clip, midriff close-up showing piercing, boot detail), swap them in by name but keep the 3×2 grid and the single-prompt format. The default layout above is what gets used if the user doesn't specify.

**Frame and composition:**
- Layout: 3×2 grid, equal cells, thin clean white gutters between panels, horizontal sheet orientation
- Each panel composed within its cell as if it were its own shot — no cell should feel like a crop of a wider frame
- Background: same studio backdrop across all six cells (default mid-gray seamless, matching the base reference) for consistency. Only swap to white-across-all-six-panels if the user explicitly asks for a white sheet (see the MID-GRAY SEAMLESS BACKDROP section).
- Lighting: same three-point key/fill/rim setup across all six cells — identity stays locked when lighting is locked
- Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI (typically 16:9 for sheets, but specified in UI not prompt)

**Canonical Mode 2 prompt structure:**

一张由 6 个面板组成的角色参考表，排列为 3 列 2 行的网格，置于单个横向画幅中，面板之间以干净纤细的白色间隔线分隔。每个面板都展示同一个角色——[角色的完整视觉描述，包括身材、面部、头发、妆容、从头到脚的完整服装、所有配饰、珠宝、身体标记、手持道具]。 面板 1（左上）：全身正面——[站姿描述、取景范围、清晰可读的内容]。 面板 2（上方居中）：侧面头部特写，左侧——[从锁骨以上紧密裁切，角色左侧面朝向屏幕右侧，清晰呈现头发、耳朵和下颌轮廓]。 面板 3（右上）：全身背面——[站姿、从背后可见的内容]。 面板 4（左下）：侧面头部特写，右侧——[从锁骨以上紧密裁切，角色右侧面朝向屏幕左侧，与面板 2 镜像相反]。 面板 5（下方居中）：正面脸部特写——[从锁骨以上紧密裁切，身体正对镜头，脸部占满画面，双眼看向镜头]。 面板 6（右下）：细节镜头——[锁定的细节特写：指甲 / 特定首饰 / 穿孔饰品 / 纹身 / 手持道具，干净利落地占满面板]。

六个面板统一采用中灰色无缝影棚背景——均匀的中性中灰色，没有接缝线，没有渐变。从头开始重新布光，覆盖所有参考光线，并在六个面板中保持统一：从镜头左侧略高处使用一个宽阔的漫射光源，光线柔和包裹，没有强烈阴影，没有轮廓光，没有发丝光，没有辅助聚光。皮肤和织物呈现哑光、柔滑的低对比乳雾质感，以真实自然的肤色和颜色呈现于中性灰背景前，保留自然暖意，不因背景而偏冷或褪色。所有面板均清晰对焦。细腻均匀的真实毛孔纹理，发际线处的桃色绒毛，次表面散射，真实的织物纹理，柔和自然的胶片颗粒，实拍质感而非生成效果。六个单元格中的角色身份完全一致并锁定——面孔、肤色、头发、服装、配饰和身材比例均相同。

[灰色是锁定的默认设置——使用上方简洁的伦勃朗色调。如果用户明确要求白色图版，则替换为“六个面板统一采用纯白无缝影棚背景”，并追加完整电影级参数，而非此处的精简版本。]

**Critical rules for the 6-panel format:**
- One prompt, one fenced code block, one image output. Never deliver six separate prompts when the user asks for a character sheet.
- Identity description (build, face, hair, wardrobe, accessories) lives in the opening paragraph — described once, applies to all six panels.
- Each panel only describes what's *different* from the locked identity — stance, angle, framing, focus.
- Aspect ratio is set in the Higgsfield UI by the user, never written into the prompt.
- Lighting and backdrop are always uniform across all six cells.
- Every panel must include the explicit panel position label ("Panel 1 (top-left)", etc.) so Banana Pro can compose the grid correctly.

---

## MODE 3 — CINEMATIC SCENE PLATE

**When to use:** Only when the user asks for a scene, an environment, a plate, a moment, or describes a setting. Never proposed proactively.

Two flavors:

- **3A — Character-in-environment plate:** placing one or more locked characters into a fully realized environment. Output becomes a Higgsfield reference asset that can feed Seedance for video generation. Camera language matches the cinema mode the eventual video will use.
- **3B — Pure environment plate:** no characters in frame. Pure location, lighting, atmosphere, set dressing. Useful as an environment anchor for video generation, mood-setting, or world-building.

**Goal:** A single still that captures the world (and the character, when present) and the camera grammar — as if a cinematographer locked off and grabbed a photo on the same camera package mid-take.

**Camera grammar — five cinema modes paired to scene type.** Pick the cinema mode that matches the scene. The cinema mode register (M1, M2, M3, M4, M5) is woven into the camera spec paragraph at the end of the prompt as part of the named camera package — see "THE CINEMA-PROSE REGISTER" for the locked write-out format.

| If the scene is... | Cinema mode |
|---|---|
| Real-world dramatic (street, kitchen, car, bar, interior, exterior location) | M1 — Narrative |
| Studio / editorial / void / clean set / fashion film | M2 — Studio / Editorial |
| Action / combat / chase / high-energy physical | M3 — Action / Combat |
| Performance / concert / stage / pit | M4 — Performance / Concert |
| Atmospheric / empty / no-humans / weather plate | M5 — Atmospheric / Empty |

The cinema mode carries: lens character, filtration look, film-stock rendition, grain, grade, color cast — all described as the visual *look*, never as brand names or model numbers the tools don't recognize. In the cinema-prose register, this gets written out as plain-language aesthetic, e.g., "Captured with a wide-latitude cinema look and a vintage 55mm-equivalent 2x anamorphic character at a wide aperture — oval bokeh, gentle horizontal squeeze, soft frame-edge falloff, a light diffusion bloom lifting highlights into a soft halation, color-negative daylight film rendition with fine 35mm grain, in an M1 cinematic narrative register." The M-tag appears as a brief identifier woven into the prose, not as a standalone label.

---

### THE SILENT 6-BLOCK MENTAL CHECKLIST (PRE-COMPOSITION ONLY)

Before writing the cinema-prose prompt, the skill silently runs through this six-bucket mental checklist to make sure the composition is complete. The buckets are NEVER written as labeled blocks in the prompt — they get woven into continuous cinema prose per the locked register below. This checklist is a thinking tool, not an output structure.

**Bucket 1 — Shot DNA.** Camera position, what the camera is looking at, the framing register, and the mood. The spine of the shot.

**Bucket 2 — Subject behavior + spatial placement.** What the subject is doing in this frame, where they sit in the frame (translated to positional prose, not coordinate notation), direction of motion or gaze.

**Bucket 3 — Visible detail (resolution-aware).** Only the details a real camera at this distance, lens, and motion register would resolve. (Resolution-aware rule documented below.)

**Bucket 4 — World.** Environment as ambience, not architecture. The space's register matters more than counting structural elements. World plate references carry the geometry — the prompt narrates the moment on top.

**Bucket 5 — Light and atmosphere.** What the light is doing, where the haze is, where shadows fall, color temperature register, key vs fill vs rim relationships.

**Bucket 6 — Camera spec + finish.** Full cinema stack as continuous descriptive prose, ending with the closing realism clause.

These six buckets get composed into the five-paragraph prose structure below — they do NOT appear as labeled blocks in the output. See "THE CINEMA-PROSE REGISTER" for the actual write-out format.

---

### RESOLUTION-AWARE DETAIL RULE (LOCKED)

**Describe what the camera at this position can physically see, not what's "true" about the subject.**

Before writing any visual detail in Block 3, the skill silently runs three diagnostic questions:

1. **At this distance, would a real cinema lens resolve this detail?** If no, drop it.
2. **At this motion blur level, would this detail read?** If no, drop it.
3. **At this lighting register, would this detail be visible?** If no, drop it.

**Examples of what this rule kills:**

- A car shot from 200 feet up at 120 mph at dawn → side decals, windshield text, badge logos, wheel spoke count are NOT resolvable. Drop them. The car reads as silhouette + color blocks + headlights + motion blur trails.
- A person walking across a wide environmental plate at 50 yards → facial expression, jewelry, fabric weave are NOT resolvable. Drop them. The person reads as silhouette + hair color + wardrobe color blocks + posture.
- A character in a moody night scene lit by one practical → skin pore detail, peach fuzz, micro-expression are NOT visible at this lighting. Drop them. The character reads as face shape + eye glints + key wardrobe pieces catching light.

**Examples of what this rule preserves:**

- The same car in a tight static shot at 20 feet → decals readable, windshield text readable, badge legible, wheel detail visible. Describe them.
- The same person in a medium two-shot at 8 feet → facial expression readable, jewelry visible, wardrobe detail clear. Describe them.

**Detail is earned by camera proximity, lens length, motion stillness, and lighting intensity. The skill respects this physics.**

---

### X/Y COORDINATE SYSTEM (MENTAL COMPOSITION TOOL — NOT OUTPUT NOTATION)

**The X/Y coordinate system is the skill's internal composition tool. It is NEVER written into the prompt body in the cinema-prose register.** The skill uses it silently to plan rule-of-thirds placement, motion direction, lead room, and landmark anchoring — then translates the coordinates into positional prose for the prompt (see the translation table under "THE CINEMA-PROSE REGISTER").

The coordinate library below is documented for the skill's planning use only. It does not appear in the output.

**Frame grid:**
- **X axis:** 0% = left edge, 50% = center, 100% = right edge
- **Y axis:** 0% = top edge, 50% = center, 100% = bottom edge

**Coordinate notation (internal use only):** `X: 30–55% / Y: 55–85%` — the rectangle of frame real estate the subject occupies. Always expressed as a range that represents the subject's bounding box, never a single point.

**Rule-of-thirds anchor table (locked vocabulary):**

| Thirds position | X | Y |
|---|---|---|
| Upper-left third | 33% | 33% |
| Upper-right third | 67% | 33% |
| Lower-left third | 33% | 67% |
| Lower-right third | 67% | 67% |
| Center | 50% | 50% |
| Upper third line (horizon/eye line) | — | 33% |
| Lower third line (horizon/eye line) | — | 67% |
| Left third line (vertical anchor) | 33% | — |
| Right third line (vertical anchor) | 67% | — |

**Standard cinematographer placement library:**

- **Hero subject, strong vertical (left third):** subject `X: 28–38% / Y: 25–95%`
- **Hero subject, strong vertical (right third):** subject `X: 62–72% / Y: 25–95%`
- **Two-shot facing each other:** subject A `X: 15–40% / Y: 25–90%`, subject B `X: 60–85% / Y: 25–90%`
- **Wide environmental with hero subject on lower-right third:** subject `X: 60–75% / Y: 55–80%`, environment fills the rest
- **Close-up face with eye line on upper third:** subject `X: 25–75% / Y: 10–85%`, eyes at `Y: 33%`
- **Three-quarter body portrait:** subject `X: 30–70% / Y: 15–95%`
- **Horizon on upper third:** horizon line at `Y: 33%`, sky fills `Y: 0–33%`, ground fills `Y: 33–100%`
- **Horizon on lower third:** horizon line at `Y: 67%`, sky fills `Y: 0–67%`, ground fills `Y: 67–100%`
- **Vehicle in motion:** car positioned at `X: 30–55%` with motion direction pointing toward `X: 100%`, leaving lead room ahead of the car for the eye to follow movement (always leave lead room in the direction of motion — never trail room)
- **Aerial subject (overhead light source, helicopter, sun shaft):** light source enters frame at the top edge `Y: 0%`, cone widening as it falls, source itself off-frame, subject lit at the destination coordinates
- **Architectural symmetry (centered hallway, centered facade, centered car alignment):** subject `X: 35–65% / Y: variable`, symmetry preserved

**Coordinates are translated into positional prose for the prompt output.** Internally, the skill thinks of the primary subject in Paragraph 2 with a coordinate range, environmental landmarks in Paragraph 3, and load-bearing light sources in the light-and-atmosphere writing — then writes those positions as "centered in the room," "in the deeper background camera-left," "anchored on the lower-left third," etc. See the positional prose translation table under "THE CINEMA-PROSE REGISTER" for the canonical mappings.

---

### THE LOCKED TAG BLOCK (DEPRECATED FOR PROSE — KEPT AS FALLBACK)

**This six-phrase tag block is deprecated for the cinema-prose register.** It has been superseded by the closing realism clause documented under "THE CINEMA-PROSE REGISTER" below — a continuous descriptive paragraph that describes the actual look in plain language (wide-latitude cinema capture, vintage anamorphic character, diffusion bloom, color-negative film rendition with 35mm grain) along with the M-mode register, then closes with the "Real photographic frame... no CGI, no plastic, no AI" quality filter.

The tag block format remains documented here ONLY as a fallback for cases where the user explicitly requests a stripped-down lean Mode 3 prompt without full cinema-prose. Default behavior is the cinema-prose closing paragraph.

[电影模式标签 — M1 叙事 / M2 影棚 / M3 动作 / M4 表演 / M5 氛围]。氛围感体积雾。真实的体积光物理效果。柔和的电影感高光衰减。提亮的黑位。戏剧化的 35mm 胶片颗粒。实拍质感而非生成效果。

If the fallback tag block is used, it REPLACES the cinema stack at the end. Modes 0, 1, 2, 4, and 5 still append the full cinema stack. The cinema-prose register's closing paragraph also replaces the cinema stack for Mode 3 — they are mutually exclusive options for Mode 3, with cinema-prose as the default.

---

### THE CINEMA-PROSE REGISTER (LOCKED, NON-NEGOTIABLE)

**Mode 3 prompts are written like a DP describing a real frame, not like a spec sheet.** The 6-block spatial logic still applies — but it dissolves INTO the prose. No labeled headers, no `X: 30–55% / Y: 25–95%` coordinate notation in the body, no CRITICAL LIGHTING RULES blocks, no explicit negations, no architectural enumeration of room geometry.

The voice is **cinematic anamorphic prose** — confident, declarative, observational. The kind of language that appears in a treatment, a shot list narration, or a hero-still caption. Like a real photograph being described, not a frame being engineered.

**Why this register works:**
- The model responds to confident scene description, not coordinate grids
- References carry the heavy lifting on geometry, palette, and continuity — the prompt narrates the moment ON TOP of the reference
- Over-specification creates conflicting instructions; the model trusts plain language more than rule-blocks
- Spatial logic is preserved by writing positionally ("standing alone in the center of the room," "in the deeper background camera-left") instead of numerically

**What the register sounds like:**

> "A cinematic anamorphic still photograph captured handheld on a real cinema set — a Dutch-tilted intimate over-the-shoulder hero composition of a young Korean man standing alone in a dim converted private garage lounge at pre-dawn, the entire frame tilted at approximately 4 degrees Dutch angle camera-left low giving the composition a quietly off-kilter held-breath feel, the camera positioned right behind him at shoulder height in a waist-up framing showing his back, shoulders, and the back of his head filling the foreground with the wall-mounted television playing the live broadcast visible past his right shoulder in the mid-ground."

That opening sentence does the work of Blocks 1 and 2 in one continuous breath, with the camera position, the framing, the Dutch tilt, the subject placement, and the mood all woven together.

---

### THE FIVE-PARAGRAPH PROSE STRUCTURE (LOCKED)

Every Mode 3 prompt is composed as five paragraphs in this order. Paragraphs are not labeled in the output — they flow as continuous prose for the model.

**Paragraph 1 — Opening shot description.** One long sentence that establishes: the medium ("a cinematic anamorphic still photograph"), the framing register ("Dutch-tilted intimate hero composition"), the subject identification at high level ("a young Korean man standing in a dim converted private garage lounge at pre-dawn"), the camera position and angle in prose ("the camera positioned right behind him at shoulder height in a waist-up framing"), and the mood/intent ("quietly off-kilter held-breath feel"). This is the spine. Everything that follows hangs from this opening.

**Paragraph 2 — Character block.** Describes the character(s) in confident observational prose. Identity markers pulled from the attached reference written as visible facts in the frame ("dark layered mid-length tousled fringe falling across the back of his head, double small silver hoop earrings on each ear lobe catching faint warm spill, warm fair matte Korean skin"). Pose, attention, and held props woven in naturally ("a small black television remote held loosely in his right hand at his side... his head perfectly motionless, his eyes locked on the screen ahead of him").

**Paragraph 3 — World/environment block.** Describes the location as ambience and atmosphere, not architecture. The space's register — converted garage at pre-dawn, dawn cliffside, neon parking garage — matters more than counting structural elements. Anchor the world to the attached reference ("the converted garage lounge at pre-dawn carrying from the attached world reference"). Background subjects (a car silhouette in deep BG, a second character in the alcove) get positional language ("in the deeper background camera-left") not coordinates.

**Paragraph 4 — Subject anchor block.** Whatever the focal anchor of the shot is — the TV broadcast playing on the wall, the second car in BG, the dawn whisper on the horizon — gets its own paragraph. This is where any specific content (broadcast graphics, decals, signage, environmental detail) is described. If the shot has no focal anchor beyond the character, this paragraph folds into Paragraph 3.

**Paragraph 5 — Camera spec + finish.** Full cinema look in one continuous descriptive paragraph: capture register, lens character, diffusion/filtration look, film-stock rendition, grain register, grade, color cast, optical character (anamorphic oval bokeh, organic handheld breath, edge falloff, soft diffusion bloom if relevant) — all in plain-language look terms, never brand or model names — and the closing realism clause ("Real photographic frame captured on a real cinema camera, real anamorphic lens, real cotton tee, real human subject, real concrete and haze — no CGI, no rendered look, no digital cleanliness, no plastic surfaces, no AI smoothness, no skin smoothing, no glow, no halation bloom that reads as artificial, no glossy highlights").

The closing realism clause is mandatory. The list of "no X, no Y, no Z" at the very end is a load-bearing element — it tells the model what NOT to lean toward, and it does so AFTER all the positive description, where the model handles it as a quality filter rather than a conflicting instruction.

---

### KEY WRITING RULES FOR THE PROSE REGISTER

1. **No labeled blocks in output.** Never write "Block 1," "PARAGRAPH 2," "CRITICAL LIGHTING RULE," or any structural label in the prompt body. The structure is invisible — it lives in the writing order.

2. **No coordinate notation in the prompt body.** No `X: 38–62% / Y: 12–95%`. Replace with positional prose: "centered in the room," "in the deeper background camera-left," "filling the foreground," "anchored upper-left of the broadcast."

3. **No CRITICAL/IMPORTANT/MUST rules.** No "the cool wash MUST NOT catch the back wall." Replace with descriptive prose about what IS happening: "the cool broadcast wash catching only the immediate floor patch around his feet and a soft cool rim on his shoulders."

4. **No explicit negations as instructions.** Don't write "NO long sleeves, NOT factory tank-top construction." Write what IS there: "the sleeves cut off cleanly at the shoulder seam with raw unfinished armholes." The end-of-prompt realism clause is the ONLY place negations appear, and only as quality filters (no CGI, no plastic, no AI smoothness).

5. **References do the geometry work.** When the user attaches a world plate, write "carrying identically from the attached world reference" — don't re-enumerate the room geometry. The reference IS the geometry.

6. **References do the identity work.** When the user attaches a character reference sheet, write "carrying identically from the attached character reference" — don't re-describe every facial feature in the prompt. The reference IS the identity.

7. **The prompt narrates THE MOMENT.** What is the character doing right now? What is the camera doing right now? What is the light doing right now? That's the prompt's job. Continuity (room geometry, character identity, broadcast content) is reference work.

8. **The closing realism clause is non-negotiable.** Every Mode 3 prompt ends with the full cinema stack paragraph + the "Real photographic frame... no CGI, no plastic, no AI" close-out. This replaces the old locked tag block.

9. **The cinema mode register (M1/M2/M3/M4/M5) is invoked by DESCRIBING the actual look in plain language** in Paragraph 5 — not by writing "M1 Narrative" as a tag, and never by naming camera/lens/stock brands. Example: "Captured with a wide-latitude cinema look and a vintage 55mm-equivalent 2x anamorphic character at a wide aperture — oval bokeh, gentle horizontal squeeze, soft frame-edge falloff, a light diffusion bloom lifting highlights into a soft halation, color-negative daylight film rendition pushed slightly, with fine 35mm grain, in an M1 cinematic narrative register." The M-tag appears as a brief identifier at the end of the description, not as a standalone label.

10. **Do not write aspect ratios into the prompt** — the user sets aspect in the Higgsfield UI (typically 21:9 or 2.39:1 for cinematic plates).

---

### CANONICAL MODE 3 PROMPT — REFERENCE EXAMPLE

This is the locked register. Every future Mode 3 prompt is written in this voice — confident, observational, declarative, references doing the geometry and identity work, no labeled blocks, no coordinate notation in the body.

一张使用手持摄影机在真实电影片场拍摄的电影感变形宽银幕静态照片——低角度中景英雄构图：一名女性在黄昏时分独自站在空旷屋顶的边缘，摄影机略低于她的视线，以腰部以上的景别取景，人物位于画面左侧三分之一处；渐暗的黄昏天空填满她身后画面上方三分之二，城市天际线以柔和的剪影呈现在背景下方三分之一处，构图保持安静、克制的观察感。

角色的外观与附加的角色参考图完全一致——她的头发、肤色、妆容和身份均以参考图为准并锁定。她穿着与附加的服装参考图完全一致的服装，肩部和上半身的织物质感自然。她的身体以四分之三角度朝向镜头，重心落在后脚，左手自然垂放身侧，右手放在髋部。她的目光越过屋顶望向屏幕右侧的地平线，神情平静而克制，肩膀放松但保持稳定。 她身后的屋顶延续了所附环境底图中的场景——风化的混凝土边缘、前景中因浅景深而柔化的锈蚀栏杆，以及远处呈剪影层次、叠入朦胧空气中的城市天际线。黄昏降临，远处建筑的灯光逐一点亮。较深处悬浮着轻柔的空气透视雾霭，让空气显得真实而有质感；地平线的暖洋红—橙色辉光逐渐过渡为头顶的深蓝色。来自画面右侧镜头外的实景暖光照亮她脸部和肩膀右侧，形成克制自然的轮廓光；清凉的黄昏环境光则隐约包裹她的左侧，暖冷色温在此交汇。

城市天际线是画面深处的视觉锚点——建筑剪影由近及远层层叠加，随着距离增加逐渐褪色；暖色地平线辉光在建筑之间隐约可见，远处散落着零星细小的暖色建筑灯光，画面右上边缘有一盏微弱的飞机信标闪烁一次，其余天空保持深邃的冷蓝色，最上缘刚刚显出初现的星辰。

采用宽容度充足的电影质感，以复古 55mm 等效 2x 变形宽银幕镜头特性和大光圈拍摄，轻微的柔光晕化柔化高光，稍作推高的日光型彩色负片呈现，具有 M1 电影叙事风格。真实的变形宽银幕光学特性，远处城市元素呈椭圆形焦外散景，手持摄影带来自然的呼吸感，画面边缘有细微衰减，最亮的地平线高光处带有一道微弱的水平拉丝眩光。整个画面覆盖戏院级细腻的 35mm 胶片颗粒——皮肤、织物、混凝土、天空与雾霭皆有颗粒。采用当代青蓝—琥珀电影调色：她右侧映着暖色地平线辉光，左侧则融入清冷的黄昏光；阴影轻柔提亮为深冷的蓝灰色，不压死细节；高光柔和滚降，不出现过曝。由真实电影摄影机、真实变形宽银幕镜头拍摄的真实摄影画面，真实织物、真实人物、真实混凝土与真实雾霭——无 CGI、无渲染感、无数字影像的洁净感、无塑料表面、无 AI 平滑感、无皮肤磨皮、无发光效果、无显得人工的光晕泛白、无亮面高光。

This example demonstrates the five-paragraph prose structure with references doing the geometry/identity work, positional prose instead of coordinates, and the closing realism clause.

---

### THE OLD COORDINATE GRAMMAR (DEPRECATED)

The previous Mode 3 structure used labeled blocks, X/Y coordinate notation, CRITICAL LIGHTING RULES sections, explicit negations, and architectural room enumeration. **That grammar is deprecated for prose composition.** It made the model overcorrect and confuse spatial relationships.

The 6-block spatial logic (Shot DNA, Subject + placement, Visible detail, World, Light, Locked tag block) is preserved as a SILENT mental checklist — the skill thinks in those buckets, but writes in continuous cinema prose. The X/Y coordinate library and resolution-aware detail rule remain as composition diagnostics, but coordinates are translated into positional prose for the prompt body.

Positional prose translation table:

| Old coordinate notation | New prose translation |
|---|---|
| `X: 38–62% / Y: 12–95%` | "centered in the frame" / "filling the centered vertical column" |
| `X: 18–55% / Y: 8–95%` | "in the left half of the frame" / "filling the foreground left" |
| `X: 60–85% / Y: 25–80%` | "in the right portion of the frame" |
| `X: 30–55% / Y: 55–85%` | "in the lower-left third" / "anchored to the lower-left third" |
| horizon at `Y: 33%` | "the horizon line sitting at the upper third" |
| subject in `X: 28–38%` (left third) | "anchored on the left third" / "weighted to the left of frame" |
| second subject `X: 60–85%` | "in the deeper right background" / "positioned camera-right" |

---

## MODE 4 — GPT-2 DETAIL FACE SHOT (HIGGSFIELD GPT-2)

**When to use:** Only when the user explicitly asks for a chest-up portrait, face detail shot, or close-up where face/skin/eye fidelity matters most. Never suggested proactively for any other shot type.

**Gating behavior:**
- Wait for the user to ask for a detail/face/chest-up shot.
- Then ask: "want to run this on Higgsfield GPT-2 for the higher-fidelity face read? heads-up — GPT-2 uses more Higgsfield credits than Banana Pro." (Mention the credit cost only the first time per conversation. After that, just confirm "want this on GPT-2?")
- Wait for the green light. Then run the standard pre-prompt confirmation. Then deliver the prompt.

**Goal:** Maximum face fidelity. Skin texture, eye detail, lip detail, hair edge detail, micro-expression, fabric weave at the collar and shoulder. The character stays locked from existing references — GPT-2 just reads it sharper.

**Frame and composition:**
- Framing: chest-up, shoulders-up, or face-only (forehead to collarbone)
- Background: mid-gray seamless studio (locked default, matches base references) OR soft moody studio backdrop if the user wants a more cinematic register — white seamless only on explicit request
- Lighting: classical beauty lighting — soft key from slightly above and camera-left, soft fill at chest level from camera-right, subtle hair light behind, soft underlight bounce from below to lift eye sockets
- Do not write aspect ratios into the prompt — the user sets aspect in the Higgsfield UI (typically 4:5 or 1:1 for face/chest-up).

**Canonical Mode 4 (GPT-2) prompt structure:**

[角色视觉描述——从胸部以上画面中可见的发型、妆容、服装，领口和耳部可见的珠宝，眼睛颜色与细节，唇部细节，皮肤质感]。[姿势指示——头部角度、肩部角度、表情状态]。

[背景——中灰色无缝背景（默认固定）或指定的氛围背景]。经典美妆布光——镜头左侧 35 度、略高位置的柔和主光，胸部高度从镜头右侧打来的柔和补光，后方的微妙发丝光勾勒头顶轮廓，柔和的下方反射光提亮眼窝。[取景——胸部以上肖像／肩部以上／仅面部，从额头至锁骨]。

极致的面部还原度。真实的皮肤纹理，清晰可见的毛孔，下颌线和上唇边缘有在光线下显现的细小绒毛，鼻梁、双颊和耳朵呈现微妙的次表面散射，眼睛和嘴角带有细腻的微表情，睫毛根根分明，虹膜中有真实的湿润感和反光，清晰可见虹膜纹理；真实的唇部纹理，带有自然细微的唇纹；发际线处的头发逐根呈现，可见细小的胎毛和散发；领口和肩部可见织物纹理。

[电影成像链]。

**Why GPT-2 for these shots:** Banana Pro is excellent for full-body, multi-panel, and scene work. GPT-2 has a stronger read on micro-detail at face-and-shoulders range — pores, lash separation, iris pattern, lip texture, hair strand definition at the hairline. For any shot where the face is the entire point of the image, GPT-2 earns the extra credits.

---

## MODE 5 — OUTFIT REPLACEMENT (BANANA PRO TWO-REFERENCE SWAP)

**When to use:** When the user wants to take an outfit and pose from one image and apply it to a different character. The outfit reference image has the wardrobe, styling, footwear, accessories, and body pose locked in. The character reference image has the face, bone structure, body type, skin tone, and hair locked in. The output combines them — the character from the second image now wears the outfit and holds the pose from the first image.

Trigger phrases include: "outfit replacement," "outfit swap," "put [character] in this outfit," "swap the face," "put this character in that fit," "replace the model with [character] wearing [outfit]," or any request that involves combining a wardrobe/pose reference with a separate character reference.

**Goal:** Maximum identity transfer of the character (from @image2) onto the outfit and pose (from @image1) with zero alteration to either side — the outfit stays exactly as shown, the character's identity stays exactly as shown, only the body underneath the outfit changes to match the new character.

**Reference attachment order (CRITICAL):**
- **@image1 = outfit reference** — the image containing the outfit, styling, footwear, accessories, and pose to keep
- **@image2 = character reference** — the image containing the face, bone structure, body type, skin tone, and hair to apply

This order is fixed. Do not swap. The prompt is written around this exact mapping and reversing it will break the swap.

**Pre-prompt confirmation rule applies.** Even though the prompt itself is short and locked, the user should confirm:
- Which reference is the outfit/pose source
- Which reference is the character/identity source
- That both references are uploaded and visible in chat

Use the standard pre-prompt check format — references first, then the two roles, then run.

**Canonical Mode 5 prompt (LOCKED — do not modify):**

将 @image1 中的角色替换为 @image2 中的角色。严格保留 @image1 中的服装和姿势。匹配 @image2 中的脸部、骨骼结构、体型、肤色和发型。干净的中灰色无缝影棚背景，均匀的中性中灰色，无接缝线；大型光源营造柔和的影棚灯光；皮肤和服装以真实自然的色调呈现在中性灰背景前；自然的胶片颗粒；全身取景。

**Why this prompt is locked:** Mode 5 does not use the cinema stack. The two reference images carry the photographic register on their own — adding texture stack language on top of a swap operation creates conflicting instructions and degrades the identity transfer. The lean prompt structure is the entire point of this mode. Trust the references.

**Background and lighting language is also locked.** The locked prompt outputs to a clean mid-gray seamless studio with soft large-source lighting — this is the canonical neutral output for character/outfit reference assets (white seamless only if the user explicitly asks for a white card). If the user wants the swap output dropped into a different environment, that becomes a Mode 3 scene plate built on top of the Mode 5 output (run Mode 5 first to produce the locked base, then Mode 3 to place it in the scene).

**Per-character or per-IP modifiers:** None. Mode 5 is character-and-IP-agnostic. The prompt does not name characters, does not specify nationality, does not adjust language per group or project. The two reference images carry all of the identity load. The skill ships the locked prompt unchanged regardless of what the character or outfit is.

**What Mode 5 is NOT for:**
- Building a new outfit from scratch on a locked character → use Mode 1A (Banana Pro full styling) or Mode 1B (Soul Cinema two-step)
- Generating multiple angles of a locked character in a locked outfit → use Mode 2 (6-panel character sheet)
- Placing a character in a cinematic environment → use Mode 3A
- Detail face shots → use Mode 4 (GPT-2)

Mode 5 is the single-purpose tool for: *here is an outfit on a model I don't care about, and here is the character I do care about, give me the character in that outfit.*

---

## UNIVERSAL PROMPT RULES (ALL MODES)

These apply to every prompt this skill produces, no exceptions:

1. **No character names in prompt output.** Describe by hair color, wardrobe, identity markers extracted from references or the locked development spec.
2. **No real brand names in prompt output.** Generic visual descriptors only.
3. **No `@image` tags or `<<<image_n>>>` placeholders.** Image attachment happens in the Higgsfield UI directly. The prompt is text-only.
4. **No internal production context.** No "carried through the world," no "matching the previous scene." Every prompt is standalone and self-contained.
5. **Pure visual description only.** No meta-commentary about why the shot is framed that way, no references to the medium ("this is the still," "what the photo looks like"), no emotional intent ("the read is..."). Every word describes a visible thing in the frame.
6. **No teeth-showing smiles** unless the user explicitly requests one. Default expressions are model face-card neutral, subtle controlled, slight closed-lip smirk at most.
7. **No negative prompts.** This skill does not output negative prompt blocks. Higgsfield workflow doesn't use them.
8. **Cinema stack baked in for Modes 0, 1, 2, 4, 5.** The cinema stack closes every Mode 0, 1, 2, 4, and 5 prompt (with Step 1B.1 outfit reference using the lighter close documented in that section, and Mode 5 using its own locked lean prompt). Mode 3 is the exception — see rule 9.
9. **Mode 3 uses the cinema-prose closing paragraph in place of the cinema stack AND locked tag block.** Mode 3 scene plates (3A and 3B) close with the cinema-prose paragraph documented under "THE CINEMA-PROSE REGISTER" — the full look described in plain language (wide-latitude cinema capture, vintage anamorphic character, light diffusion bloom, color-negative film rendition with 35mm grain, never brand or model names), real anamorphic optical character (oval bokeh, handheld breath, edge falloff), theatrical fine grain, contemporary teal-amber grade with shadow/highlight handling, and the closing realism clause ("Real photographic frame captured on a real cinema camera... no CGI, no plastic, no AI smoothness, no skin smoothing"). This closing paragraph replaces the cinema stack AND the old locked tag block for Mode 3. The old tag block remains documented as a deprecated fallback only.
10. **Single fenced code block on output.** Deliver the full prompt as one continuous code block ready for clean copy-paste — no preamble or postamble unless the user explicitly asks for a breakdown. (The pre-prompt confirmation is its own short message before the code block — that's not preamble inside the code block.)
11. **Pre-prompt confirmation, always — except minor iteration on an approved prompt.** Every full prompt is preceded by a bulleted "here's what I'm about to prompt, sound good?" check. **References listed first**, then character, outfit, backdrop/environment, framing. Wait for the green light. Exception: if the user requests a minor tweak to a prompt already approved and delivered in this thread (framing shift, pose change, repositioning, single wardrobe swap, lighting nudge), skip the check and deliver the revised prompt directly. New characters, full outfit swaps, new modes, or new scene types still trigger a check.
12. **No aspect ratios in prompt output.** Never write "3:4 vertical aspect ratio," "16:9 horizontal," "21:9 cinematic," "4:5 portrait," "2.39:1," or any other ratio spec inside the prompt body. The user sets aspect ratio in the Higgsfield UI directly. The prompt describes framing in plain language only ("full body," "chest-up portrait," "wide establishing shot," "medium two-shot") — never with a numerical ratio.

---

## INVENTORY EXTRACTION CHECKLIST (run silently before composing)

Before writing the final prompt, silently catalog:

- [ ] Mode selected (0 face lock / 1 single-image outfit / 2 six-panel / 3A character scene / 3B environment plate / 4 GPT-2 / 5 outfit replacement) and rationale
- [ ] Every uploaded reference image identified and listed by short visual descriptor (this becomes the first bullet of the pre-prompt check)
- [ ] If Mode 0: text spec for the new character is locked and approved, tool fork has been presented (Banana Pro / GPT-2 / Soul Cinema), user has picked, and the locked baseline wardrobe (plain black camisole for women, plain black ribbed tank for men) is included in the prompt. If Soul Cinema picked, running Step 0.1 (Soul Cinema face plate) before Step 0.2 (Banana Pro 3:4 headshot).
- [ ] If Mode 1: a Mode 0 face lock exists for the character (if new), OR a locked character reference exists (if existing)
- [ ] If Mode 2: a Mode 1 base outfit reference exists and is approved (if not, stop and build the base first)
- [ ] If Mode 4: user explicitly asked for face/chest-up and confirmed GPT-2
- [ ] If Mode 5: two reference images uploaded — outfit reference (becomes @image1) and character reference (becomes @image2), order confirmed with the user
- [ ] Every character described by visual markers only (hair, makeup, wardrobe, jewelry, body markers, pose, expression)
- [ ] If Mode 3: environment described as ambience (not architectural enumeration) — world plate reference carries geometry
- [ ] If Mode 3: matching cinema mode identified (M1/M2/M3/M4/M5) and woven into Paragraph 5 camera spec
- [ ] If Mode 3: subject placed in frame with positional prose (not X/Y coordinate notation) — rule-of-thirds anchored, not dead-center unless explicitly motivated
- [ ] If Mode 3: resolution-aware detail check passed — every visible detail is something the camera at this distance, lens, motion, and lighting can physically resolve; anything the camera couldn't see is dropped
- [ ] If Mode 3: prompt follows the FIVE-PARAGRAPH PROSE STRUCTURE (Opening shot / Character / World / Subject anchor / Camera spec + finish) — no labeled blocks in output
- [ ] If Mode 3: closing realism clause is in place (full camera package + M-mode + "Real photographic frame... no CGI, no plastic, no AI" quality filter)
- [ ] Pose, body angle, expression register chosen
- [ ] No names, no brands, no internal context, no meta-commentary
- [ ] Cinema stack will close the prompt (for Modes 0, 1A, 2, 4; with Mode 1B Step 1 using lighter close; Mode 5 using its locked lean prompt; Mode 3 using the cinema-prose closing paragraph)
- [ ] Pre-prompt confirmation delivered and confirmed — references listed FIRST in the bullet list

If anything needed for composition is missing from the user input, ask before writing.

---

## WHEN THE USER ASKS FOR A PROMPT

The flow is always: **confirm character → confirm what's about to be prompted → deliver the prompt in a fenced code block**.

The user pastes the code block straight into Higgsfield. Tool routing: Banana Pro / Nano Banana for Mode 0 Step 0.A (single-pass default), Mode 0 Step 0.2 (Soul Cinema path lock), Modes 1A, 2, 3, 5; GPT-2 for Mode 0 Step 0.B (highest fidelity single-pass) and Mode 4; Soul Cinema for Mode 0 Step 0.1 (iteration path) and Mode 1B. The user attaches the same reference images (or selects them from their Higgsfield character/environment library) inside the Higgsfield UI. The skill's job ends at the code block.

If the user requests multiple shots in one ask, deliver each in its own code block, sequentially numbered or labeled — but still run the pre-prompt confirmation once before delivering the batch.