把一张人像照变成动起来，是 AI 视频里情感价值最高的玩法——会呼吸的婚纱照、又笑了的爷爷奶奶、摇尾巴的宠物。这也是最容易把人脸搞崩的环节，因为只要动作一大，面部就重排。诀窍是克制：一个微动作、机位锁死、5–8 秒以内。下面 10 个可复制的人像转视频模板，都是为了把原本的脸保留下来调出来的。
任何人像转视频都该锁住的 5 个要素
上传原图之前先确认：
- 镜头（Lens）——继承原图，不要再描述（在 24mm 人像上写
85mm会引入畸变） - 光线（Light）——同样继承；写
lighting unchanged from source - 运动（Motion）——只做一个微动作：眨眼、微笑、轻偏头、发丝飘、面纱抚动
- 色调（Palette）——
colors unchanged from source几乎都是对的 - 5–8 秒克制——超过 6 秒人脸就开始飘
铁律：一个微动作 + 机位锁死。再加第二个动作或者任何机位运动，脸就变形。
10 个可复制的 Prompt 模板
1. 婚纱照·面纱微动
The bride in the frame stays still. Her veil sways gently as if a light breeze passes through. A single soft strand of hair near her temple moves slightly. No facial movement. Camera is locked, no movement at all. Lighting and colors unchanged from source. Duration: 5 seconds.2. 全家福·浅笑慢推
Each person in the family photo gives a very small, soft natural smile, holding for the full clip. No head movement, no blink. Camera does a barely perceptible slow zoom in, about 5 percent total. Lighting unchanged. Duration: 6 seconds.3. 老人宠物照·摇尾巴
The dog in the photo wags its tail gently, side to side, two or three times across the clip. Dog's body, head, and face stay completely still. Owner stays still. Camera is locked. Lighting and colors unchanged. Duration: 5 seconds.4. 祖辈老照片·轻摇镜
The subjects in the vintage photo remain completely still. Camera does a very slow gentle pan from left to right across the frame, rotating about 5 degrees total. Subtle film grain breathes. Sepia tone and lighting unchanged. Duration: 6 seconds.5. 复古写真·眨眼配甩发
The model in the pinup photo gives a single playful wink with her right eye, then a small hair flip with her right hand near her shoulder. Single sequence, no repeat. Camera is locked. Lighting and colors unchanged. Duration: 5 seconds.6. 网红肖像·甩发加微笑
The subject in the portrait gives a soft smile and a small head tilt to her right, with a single strand of hair brushing past her cheek. No eye movement other than blink. Camera locked, no zoom. Lighting and palette unchanged. Duration: 5 seconds.7. 商务证件照·轻挑眉
The subject in the corporate headshot raises both eyebrows very slightly once, holding the position for the rest of the clip. No smile, no head movement, no other facial change. Camera locked. Lighting unchanged. Duration: 5 seconds.8. 学生照·轻笑
The student in the school photo gives a small natural laugh, with very small shoulder movement and a single soft eye-crinkle. No head turn, no zoom. Camera is locked. Lighting and uniform colors unchanged. Duration: 5 seconds.9. 时尚封面·缓转头
The model on the fashion cover turns her head slowly from facing camera to a three-quarter profile to her left, ending in a soft confident expression. Body and shoulders stay still. Camera locked. Lighting and palette unchanged. Duration: 6 seconds.10. 宠物肖像·眨眼加耳动
The cat in the pet portrait blinks once slowly, then one ear twitches subtly. Rest of the body stays still. Camera is locked. Lighting and colors unchanged. Duration: 5 seconds.5 个常见误区
- 同时要微笑、转头、眨眼——只挑一个微动作
- 写新的光线描述——会覆盖原图、把脸的光面也重画
- 给人像加机位运动——哪怕一点点摇都会破坏面部一致性
- 不写
lighting unchanged from source——模型按默认调色，肤色就漂了 - 特写超过 6 秒——脸开始可见地变形
5 个进阶用法
- 同一张图跑三次——成本最低的方差控制，挑最干净的那条
- Veo 给全家福加一行环境音——
ambient: soft laughter, distant chatter - 接续三段人像片——同一个人，每段不同微动作
- 用慢叠化串起来——把”静变动”的魔法感卖足
- 先把原图放大——低于 1024px 的源 5 秒内身份就崩
Sora / Veo / Kling 在人像上的分工
各家身份保留差别极大：
- Kling——较长单镜下身份保留最稳。适合婚纱和全家福。东亚面孔特别在行。
- Runway Gen-3——西方面孔和自然微表情强。网红和时尚照默认选它。
- Veo 3——同步微表情加一行轻环境音（
ambient: soft laughter）最强。家庭片想加一点声音用它。 - Sora——更风格化；适合复古或电影感、可以接受一点诠释的画面。身份漂移可能偏高。
按需求选：身份要 100% 不变 → Kling；要声音 → Veo；要风格化 → Sora。
不同情绪下的人像配方
- 浪漫 / 婚纱——面纱抚动 + 机位锁死 + 5 秒 + 暖色调不变
- 追忆 / 怀旧——轻摇镜 + 黄褐色或低饱和色调 + 6 秒 + 不做面部动作
- 俏皮 / 社交——眨眼或甩发 + 机位锁死 + 5 秒 + 饱和色调
- 职业 / 商务——挑眉 + 机位锁死 + 5 秒 + 中性冷调
- 温情 / 家庭——浅笑 + 几乎察觉不到的微推 + 6 秒 + 暖色调
FAQ
Q：为什么主体的脸中途变形？ A：要么片子太长（剪到 5 秒），要么动作太大（换更小的），要么原图分辨率太低（放大到 1024px+）。
Q：哪家模型做人像转视频最好？ A：Kling 和 Runway Gen-3 在身份保留上领先。Veo 的微表情细节强。Sora 偏风格化；同一图三家都跑。
Q：怎么把已故亲人的照片得体地做动？ A：这里的模板就是按”克制”调的。只做一个微动作（微笑、眨眼、轻摇镜），不要转头、不要说话。
Q：怎么把原图的颜色完全保留？
A：永远写 lighting and colors unchanged from source。不写这句模型就会漂回它默认色调。
Q：人像视频最适合什么画幅？ A：跟原图严格一致。强行换比例就会裁脸或者扭脸。