AI 视频角色一致性：让每个镜头都是同一张脸的 9 条多镜头提示词

做 AI 短剧和连载内容的人有一个共同的噩梦：同一个角色，换个镜头就换了张脸。角色一致性是剧情类 AI 视频的头号难题——但解法不是某一个”神功能”，而是三层纪律的叠加。下面先讲清楚”为什么会漂”，再给三层解决方案、9 条”多镜头同角色”的完整提示词（每组 2–3 个镜头共用同一段角色描述），以及可以永久复用的角色描述固定字段模板。示例适用于 Veo 3.1、Kling 4.0、Seedance 2.5、Hailuo 3 等主流模型。

为什么角色的脸总会漂

• 每条视频是独立采样的。 模型每次生成都从零”摇”一遍”黑发年轻女性”。文字描述不足以锁定一张脸，每次摇出来的结果都会偏一点。
• 你在镜头之间改了措辞。 把”long black hair”改成”dark hair”，等于递给了采样器另一个人。
• 角度和光线最先暴露漂移。 侧脸稳住了，特写可能就融化。
• 远景里的脸像素太少。 等你切进特写，模型就重新发明了一张脸。

公式：一段锁死的角色描述块

text [CHARACTER LOCK — copy this block into every shot prompt, unchanged] A 28-year-old East Asian woman, oval face, almond-shaped dark brown eyes, small mole under her left eye, shoulder-length straight black hair with a side-swept fringe, cream trench coat over a white turtleneck, silver watch on her left wrist.

锁死描述块的三条规则：

1. 逐字粘贴。 每个镜头都用一模一样的词、一样的顺序，不改写、不缩写。
2. 锁定 2–3 个”扛得住任何角度”的锚点特征——泪痣、疤痕、眼镜、精确发色、一件标志性衣服。独特比漂亮更重要：脸会漂，左眼下的痣会跟着走。
3. 先身份，后动作。 角色块放最前面，然后才是动作、运镜、灯光。

第二层：参考图功能（已核实的模型功能）

文字能锁住的身份有限。目前主流模型的参考图系统如下：

• Veo 3.1 — Ingredients to Video： 每次生成最多 3 张参考图（角色、物体、场景/风格），用于跨镜头保持角色一致。
• Kling 4.0 — Omni Reference： 每次请求最多 15 个多模态参考（图片 + 最多 5 段视频片段），面向角色与风格的连续性。
• Seedance 2.5 — 混合参考集： 单次生成最多 30 张图片、10 段视频、10 条音频，提示词中以 [Image1]、[Vid1] 引用；官方建议每次生成都用同一张身份参考图。
• Hailuo 3（MiniMax H3）— Omni-Reference： 最多 9 张参考图外加视频、音频参考来钉住跨镜头身份。

两条规则：每次都上传同一套身份参考图（一张面部特写 + 一张全身着装图）；用原创角色——Veo 等模型会拦截真实人物和名人肖像。

第三层：用镜头设计绕开漂移

• 动作在一个镜头内完成——切镜越少，重摇越少。
• 转场时切走脸： 手部、物件、背影、剪影、过肩。
• 整组镜头统一视线方向、主光方向和景别距离。
• 先生成主特写，让它成为其余镜头的锚点参考。

9 条”多镜头同角色”复制提示词

每组的 2–3 个镜头逐字复用同一段角色描述块。复制镜头时把角色块一起带走。

A 组 — Maya：雨夜小巷（3 镜）

A1. 全景建立

text Character (identical in every shot): a 28-year-old East Asian woman, oval face, almond-shaped dark brown eyes, small mole under her left eye, shoulder- length straight black hair with a side-swept fringe, cream trench coat over a white turtleneck, silver watch on her left wrist. Wide shot: she walks down a rain-soaked alley at night, stops under a red paper lantern and looks up at the rain, neon reflections on wet asphalt, slow push-in, cinematic 35mm look, gentle rain audio, bittersweet mood.

点评：短剧开场建立镜头；建议第一个生成，之后当参考图锚点。

A2. 特写反应

text Character (identical in every shot): a 28-year-old East Asian woman, oval face, almond-shaped dark brown eyes, small mole under her left eye, shoulder- length straight black hair with a side-swept fringe, cream trench coat over a white turtleneck, silver watch on her left wrist. Medium close-up: she stands under the red paper lantern, takes out her phone, reads the screen and smiles softly, warm lantern key light, cool neon rim light, shallow depth of field, subtle handheld drift, soft rain audio, quiet hopeful mood.

点评：情绪镜头；主光方向（灯笼暖光）与 A1 保持一致是关键。

A3. 过肩出门

text Character (identical in every shot): a 28-year-old East Asian woman, oval face, almond-shaped dark brown eyes, small mole under her left eye, shoulder- length straight black hair with a side-swept fringe, cream trench coat over a white turtleneck, silver watch on her left wrist. Over-the-shoulder shot from behind her as she pushes open a noodle-shop door, a bell chimes, warm steam and light spill across her face, she glances back at the alley, camera follows two steps inside, handheld, door chime and kitchen ambience audio.

点评：过肩 + 回眸是”转场不切脸”的范本写法，正脸只占一小段时间。

B 组 — Daniel：夜班侦探（2 镜）

B1. 钠灯下的证物

text Character (identical in every shot): a 40-year-old Black man, shaved head, strong jawline, dark brown eyes, short greying beard, thin scar through his right eyebrow, tall athletic build, charcoal wool overcoat over a navy shirt, black leather gloves. Interior medium shot in a parked car at night: he opens a manila evidence envelope under a warm sodium parking-garage light, photographs slide onto his gloved palm, slow push-in from the passenger side, orange lamp glow through the windshield, tense mood, muffled rain audio.

点评：悬疑类质感镜头；眉间疤痕是本角色锚点，任何镜头都别省。

B2. 走向门口

text Character (identical in every shot): a 40-year-old Black man, shaved head, strong jawline, dark brown eyes, short greying beard, thin scar through his right eyebrow, tall athletic build, charcoal wool overcoat over a navy shirt, black leather gloves. Wide shot: he steps out of the car, closes the door and walks toward a brick entrance, silhouetted against warm doorway light, static low-angle camera, wet street reflections, cold blue night grade, footsteps audio.

点评：剪影镜头几乎不考验五官一致性，是两段正脸戏之间最安全的过渡。

C 组 — Sofia：厨房带货主持（2 镜）

C1. 咖啡机演示

text Character (identical in every shot): a 32-year-old Latina woman, warm olive skin, round face, curly shoulder-length dark brown hair tied half-up, hazel eyes, gold hoop earrings, sage green linen apron over a white tee. Medium shot in a sunlit kitchen: she places a cup under a stainless espresso machine, pulls the lever, espresso streams out and steam curls upward, slow push-in over the counter, bright morning window light, cheerful mood, brewing sounds audio.

点评：带货/种草类的标准演示镜头，金耳环 + 半扎卷发 + 围裙是三重锚点。

C2. 第一口

text Character (identical in every shot): a 32-year-old Latina woman, warm olive skin, round face, curly shoulder-length dark brown hair tied half-up, hazel eyes, gold hoop earrings, sage green linen apron over a white tee. Close-up: she lifts the cup, inhales the aroma, closes her eyes and smiles, soft morning light across her face, shallow depth of field, static camera, warm inviting mood, kitchen ambience audio.

点评：与 C1 同一场景同一光线，成组生成时一致性最稳。

D 组 — Leo：晨跑男孩（2 镜）

D1. 大桥跑步

text Character (identical in every shot): a 24-year-old white man, wavy sandy-blond hair, blue eyes, light stubble, lean runner's build, navy running jacket over a grey tee, black shorts, blue wireless earbuds. Wide shot: he jogs across a suspension bridge at sunrise, camera tracking alongside at matching pace, long shadows, warm golden backlight over the river, breath vapor in cool air, energetic calm mood, footsteps and wind audio.

点评：运动生活方式素材；蓝色耳机是远景里也能认出的锚点配饰。

D2. 剪影放松

text Character (identical in every shot): a 24-year-old white man, wavy sandy-blond hair, blue eyes, light stubble, lean runner's build, navy running jacket over a grey tee, black shorts, blue wireless earbuds. He stops at a park fountain, drinks and catches his breath, backlit by low sun so his face reads mostly in silhouette, camera arcs behind his shoulder toward the skyline, warm rim light, reflective mood.

点评：逆光剪影 + 过肩弧形运镜，用”看不清的脸”结束段落，规避特写漂移。

角色描述固定字段模板

给系列里的每个角色都填一份。把散文版（像上面的锁死块）粘进提示词，用字段表检查有没有漏项。

text CHARACTER LOCK (copy into every shot; keep field order and wording identical): AGE + ETHNICITY: FACE SHAPE / EYES / NOSE / BROWS / MOUTH: DISTINGUISHING MARKS (must survive any angle — mole, scar, freckles, glasses): HAIR (color, length, texture, style): BUILD + HEIGHT IMPRESSION: WARDROBE (exact garments, colors, layers): ACCESSORIES (with side — watch on left wrist, ring on right hand): VOICE (only if dialogue/audio matters):

3 个额外的一致性技巧

1. 给参考图文件做版本号（CHAR_MAYA_v1_face.png），系列中途绝不混用版本。
2. 每条提示词统一面部主光方向，光影几何一致，人就”是同一个人”。
3. 拍摄单元保持短。 用 15–20 秒的叙事单元拼系列，窗口越短，身份越稳。

FAQ

提示词一字不改，为什么脸还是漂？ 每次生成独立采样。解法：锚点特征 + 锁死描述块 + 每次生成都挂同一套参考图。

哪个模型做角色一致性最强？ 四家都有参考系统：Kling 4.0 Omni Reference（15 个参考）、Seedance 2.5（大规模混合参考集）、Veo 3.1 Ingredients（3 张图）、Hailuo 3 Omni-Reference（9 张图）。按你管线的其余环节选：运动风格、音频、片长。

能用真人演员或网红的照片吗？ 不能。Veo 等模型会拦截真实人物和名人肖像。用生成肖像 + 锁死描述块做原创角色。

是一条多镜头长视频好，还是分条生成再剪？ 单条多镜头生成天然保身份，但片长有上限；分条生成必须靠锁死块 + 同一套参考图——正是这套提示词的工作流。

首尾帧控制有帮助吗？ 有。图生视频钉住首帧和尾帧（Seedance 2.5、Hailuo 3 等支持），锚定镜头的开始与结束。

角色建一次，锁死块用到老。系列成片之后，FLOWNIB 可以帮你把做好的剧集一次铺开到多个平台。