# seeapi-zombie prompts

只有 ```text 块会发出去。和 `exp/zombie/prompts.md` 用同一套规则——输入用自然语言指代,只增不改。这里直接复用 `exp/zombie/prompts.md` 的 **v3**(本来就是为"参考视频 + 2 张照片"这个用法写的,原样照抄,没有改字)。

照片顺序假设(见 `exp.json` 的 `image_order_assumption`):第 1 张 = cho(幸存者/举枪的人),第 2 张 = dean(变僵尸的人)。原视频里举枪的是女性,所以这么配;如果想反过来(dean 举枪、cho 变僵尸),告诉我换。

## v3 — couple,用参考视频(来自 exp/zombie/prompts.md,未改动)

```text
Use the reference video as the exact shot-by-shot template: keep its camera angles, cuts, timing, movements, the wooden cabin and the golden wheat field at sunset. Replace the woman who holds the pistol with the person in the first reference image, and replace the zombie with the person in the second reference image. Take each person's face, hair, skin tone, apparent age and clothing from their reference image; in the wheat-field scene the second person appears healthy again, in light clothes. Keep the zombie look from the reference video (pale cracked skin, clouded white eyes) without adding any gore. The pistol is never fired. Only these two people appear.
```

## v4 — couple,身份锁定加强版(2026-10-30,v3 跑完后发现问题:僵尸那边换成功了,幸存者那边的脸基本没换,只换了衣服)

### 根因分析(有真实生成结果做对照)

逐帧截图对比(0s/5s/8s/12s)+ 两张人物参考照:

- **僵尸(第 2 张参考图 = dean)**:8s 扑过去那镜头能看清肤色、发型、那件深绿夹克,确实换成了 dean。**换得对。**
- **幸存者(第 1 张参考图 = cho)**:从 0s 到 8s,脸型、发色(金棕色波浪长发)、五官跟参考视频原片里的演员几乎一模一样,完全不是 cho(东亚女生、黑直发)的样子,只有外套颜色变了。**没换对。**

原因(查了外部资料,见下方"调研依据"):

1. **模型本身的倾向**:参考视频里幸存者那段大量是哭泣的极近特写(表情细节多、动态强),这类镜头模型对原片人脸的"连续性"依赖更强,用静态照片去覆盖一张正在大哭的动态脸,比覆盖一张本来就要大改(僵尸化)的脸难得多。研究资料里管这个叫 identity leakage——模型会从目标视频本身的脸部信息里"偷"身份特征。
2. **prompt 写法不够:v3 没有用文字描述 cho/dean 具体长什么样**,只写了"the person in the first/second reference image",纯靠图片参考。对比 `exp/seeapi-hotel-lobby`(文字里写了"platinum hair, green eyes"等具体特征,身份换得很准),差别很明显。

### 调研依据

- Scenario.com 的 MiniMax Hailuo(H3)指南:prompt 里**先用"Use Image 1 as..."/"Use Image 2 as..."显式分配角色**,比不带编号的引用更可靠;要让嘴型对上台词,得**把台词放在引号里并标注是谁说的**("X says: '...'"),单纯叙述"某人在说话"模型不会正确对嘴型。
- Runware 的 Reference-driven consistency 指南:多参考图场景下,**按编号("image 1"/"image 2")指代、角色分配要明确**。
- 人脸替换相关论文(如 identity leakage 现象):换脸时模型容易从目标视频本身偷取身份信息,解决办法是在文字里反复、具体地描述"要保留"的那个身份特征,压过模型对原片的依赖。

### v4 prompt(改动:开头先显式分配角色+具体长相描述,每个人脸特写镜头都重复点名身份特征,强调"覆盖原片演员的脸"而不是含糊的"take their face from the reference image")

```text
Use the first reference image as the survivor: a young East Asian woman with long straight black hair and brown eyes. Use the second reference image as the zombie: a young Black man with short dark curly hair and a light mustache. These two identities must replace the original actors' faces completely in every single shot, including the extreme close-ups — keep her East Asian features and black hair, and his dark skin and curly hair, visually consistent with their reference images throughout, overriding the original video's faces.

Use the reference video as the exact shot-by-shot template: keep its camera angles, cuts, timing, movements, the wooden cabin and the golden wheat field at sunset. Segment 1 (0-4 s): an extreme close-up of the East Asian survivor's face crying, the pistol barrel blurred in the foreground — her black hair and East Asian features stay clearly visible; then an over-the-shoulder shot from behind her as the dark-skinned, curly-haired zombie walks slowly toward her out of the backlit doorway; then a medium shot of her aiming with both hands, tears running. Segment 2 (4-6 s): two extreme close-ups of the zombie's face — pale cracked skin and clouded white eyes layered onto the Black man's own face and short curly hair, not the original actor's face. Segment 3 (6-10 s): a close-up of her tearful eyes, still clearly her East Asian face; a medium shot as she lowers the pistol, drops it and opens both arms; a wide shot as the zombie lunges forward and leaps toward her, the pistol lying on the floor. Segment 4 (10-15 s): a hard cut to the golden wheat field at sunset — the same East Asian woman and the same Black man, now healthy and in light clothes, hug and spin, laughing and crying; finish on an extreme close-up of their faces cheek to cheek, both faces clearly matching their reference images. Keep the zombie look from the reference video (pale cracked skin, clouded white eyes) without adding any gore. The pistol is never fired. Only these two people appear.
```

### v4 prompt(中文翻译,仅供你确认用,实际发送的是上面的英文)

```text
把第一张参考图当作幸存者:一个有着黑色长直发、棕色眼睛的年轻东亚女生。把第二张参考图当作僵尸:一个留着深色短卷发、留着淡淡胡须的年轻黑人男生。这两个身份必须在每一个镜头里(包括极近特写)完全替换掉原演员的脸——她的东亚五官和黑发、他的深色皮肤和卷发,要在全片保持和参考图一致,压过原视频里的脸。

用参考视频做逐镜头模板:保留它的运镜、剪辑点、节奏、动作,木屋和夕阳麦田的场景。第 1 段(0-4秒):幸存者(东亚女生)哭泣的极近特写,枪口虚化在前景——她的黑发和东亚五官要清晰可见;然后是她背后的过肩镜头,深色皮肤、卷发的僵尸从逆光门口慢慢走来;然后中景,她双手持枪,泪流满面。第 2 段(4-6秒):僵尸脸的两个极近特写——苍白开裂的皮肤、发白的眼睛叠加在那个黑人男生自己的脸和短卷发上,不是原演员的脸。第 3 段(6-10秒):她含泪眼睛的特写,依然清楚是她的东亚脸;中景她放下枪、扔掉枪、张开双臂;全景僵尸扑向她,枪躺在地上。第 4 段(10-15秒):硬切到夕阳麦田——同一个东亚女生和同一个黑人男生,已经恢复健康、穿着浅色衣服,拥抱旋转;结尾是两人脸贴脸的极近特写,两张脸都要清楚对应参考图。保留参考视频里僵尸的样子(苍白开裂皮肤、发白眼睛),不要加血腥元素。枪始终没有开火。全程只有这两个人出现。
```

还没提交,等你确认。