# 多人换人提示词（红模特主角 + 6 个白模特配角，7 张三视图）

- 生成时间：2026-09-14
- 场景：源视频 = `<Video 1>`，画面里红模特是主角、白模特是配角（按左右顺序 1–6 号）
- 参考图 = 7 张三视图（每张 = 1 个角色，正/侧/背三视角）
- 校验：`python3 scripts/validate_h3_ref2va_prompt.py --profile identity-migration <文件>` → **OK: valid H3 Ref2VA prompt (identity-migration)**

## 一、编号映射（必须与 Director 里 参考图1..7 的顺序完全一致）

| 参考图槽位 | 提示词标签 | 目标角色 | 替换源视频里的谁 |
|---|---|---|---|
| 参考图 1 | `<Picture 1>` | `<Subject 2>` | `<Subject 1>` 红模特（主角） |
| 参考图 2 | `<Picture 2>` | `<Subject 4>` | `<Subject 3>` 白模特 1（左起第一个） |
| 参考图 3 | `<Picture 3>` | `<Subject 6>` | `<Subject 5>` 白模特 2 |
| 参考图 4 | `<Picture 4>` | `<Subject 8>` | `<Subject 7>` 白模特 3 |
| 参考图 5 | `<Picture 5>` | `<Subject 10>` | `<Subject 9>` 白模特 4 |
| 参考图 6 | `<Picture 6>` | `<Subject 12>` | `<Subject 11>` 白模特 5 |
| 参考图 7 | `<Picture 7>` | `<Subject 14>` | `<Subject 13>` 白模特 6 |

- 源实体用奇数号（1/3/5/7/9/11/13），目标身份用偶数号（2/4/6/8/10/12/14）；`<Video 1>` 永远是那个参考视频。
- 若视频里白模特不是 6 个：**成对增删**（`<Subject 3>`+`<Picture 2>` 是一对），并保证 Subject / Picture 编号**从 1 起连续**、每段都要在 `subject_definitions:` 里以 `<Subject N> is ...` 形式定义，否则校验器报错。

## 二、五条铁律

1. **三视图 = 1 个人，不是 3 个人**。7 张三视图若不声明，H3 会把画面渲染成 21 个（或至少乱复制）。提示词里 `reference_interpretation:` 一节专门干这个，删不得。
2. **提示词标 `<Picture N>`，不要写 `<Image N>`**；也**不要**用 `@asset` 这类派生引用别名（那是 studio 编排路径的写法，Director 里会报 `unstable_reference_tag`）。
3. **不要**对源模特用 `attribute_transfer` / `partially_preserved` / `weak_reference`——那会削弱替换，出现「原人物没换掉」。
4. **必须靠 `<Video 1>` 传动作**（RV2V），不要靠单帧姿态；`--mode r2v` 不接受源视频。
5. **分段**：单段上限 362 帧（15.083s）。源视频 36.2s / 1086 帧 @30fps → 至少 3 段；每段用**同一份提示词**、Picture 编号不变（global 编辑模式下所有段共用一份全局提示词）。提示词里我只写了 `[Shot 1]`（无时间戳），就是为了分段复用不冲突；若你自己切片并确知镜头边界，再补 `[Shot N] At MM:SS.mmm,`。

## 三、生成参数（这次已验证）

- 采样：4 步 + `minimax_h3_ref2v_turbo_4step_v0.1`（strength 1.0）、`res_multistep` / `simple`、shift 视频 12 / 音频 3、cfg 1.0
- 分辨率：竖版 480×864 或 16:9 864×480（0.4MP）
- 采样步数要填 **4**（不是 25）

## 四、提示词全文

```text
subject_definitions:
<Video 1> is the source reference video for the target video edit. It supplies the original performance, body movement, pose timing, camera movement, framing, background, lighting, spatial relationships, group formation, and interactions, but it supplies no identity appearance.
<Subject 1> is the original protagonist in <Video 1>: the model wearing the red outfit, the leading figure of the group. <Subject 1> is a source entity to be completely replaced, and its identity must not survive in the target video.
<Subject 2> is the target protagonist, whose identity and complete visual appearance come from <Picture 1>.
<Subject 3> is the first white-outfit supporting model in <Video 1>, first in the left-to-right order of the white models beside the protagonist. <Subject 3> is a source entity to be completely replaced.
<Subject 4> is the target supporting character whose identity and complete visual appearance come from <Picture 2>. <Subject 4> replaces <Subject 3> in the same position and the same role.
<Subject 5> is the second white-outfit supporting model in <Video 1>, second in the left-to-right order. <Subject 5> is a source entity to be completely replaced.
<Subject 6> is the target supporting character whose identity and complete visual appearance come from <Picture 3>. <Subject 6> replaces <Subject 5> in the same position and the same role.
<Subject 7> is the third white-outfit supporting model in <Video 1>, third in the left-to-right order. <Subject 7> is a source entity to be completely replaced.
<Subject 8> is the target supporting character whose identity and complete visual appearance come from <Picture 4>. <Subject 8> replaces <Subject 7> in the same position and the same role.
<Subject 9> is the fourth white-outfit supporting model in <Video 1>, fourth in the left-to-right order. <Subject 9> is a source entity to be completely replaced.
<Subject 10> is the target supporting character whose identity and complete visual appearance come from <Picture 5>. <Subject 10> replaces <Subject 9> in the same position and the same role.
<Subject 11> is the fifth white-outfit supporting model in <Video 1>, fifth in the left-to-right order. <Subject 11> is a source entity to be completely replaced.
<Subject 12> is the target supporting character whose identity and complete visual appearance come from <Picture 6>. <Subject 12> replaces <Subject 11> in the same position and the same role.
<Subject 13> is the sixth white-outfit supporting model in <Video 1>, sixth in the left-to-right order. <Subject 13> is a source entity to be completely replaced.
<Subject 14> is the target supporting character whose identity and complete visual appearance come from <Picture 7>. <Subject 14> replaces <Subject 13> in the same position and the same role.
<Picture 1> is the character turnaround sheet of the target protagonist <Subject 2>: one single person shown from the front, side, and back views.
<Picture 2> is the character turnaround sheet of the target supporting character <Subject 4>: one single person shown from the front, side, and back views.
<Picture 3> is the character turnaround sheet of the target supporting character <Subject 6>: one single person shown from the front, side, and back views.
<Picture 4> is the character turnaround sheet of the target supporting character <Subject 8>: one single person shown from the front, side, and back views.
<Picture 5> is the character turnaround sheet of the target supporting character <Subject 10>: one single person shown from the front, side, and back views.
<Picture 6> is the character turnaround sheet of the target supporting character <Subject 12>: one single person shown from the front, side, and back views.
<Picture 7> is the character turnaround sheet of the target supporting character <Subject 14>: one single person shown from the front, side, and back views.
<Picture 1> through <Picture 7> are seven different characters, exactly one single person per sheet.

summary:
[video editing + character replacement] three-view identity reference
The target video is an edited version of <Video 1>.
Completely replace <Subject 1> with <Subject 2>, whose identity and complete visual appearance are defined by <Picture 1>. In the same way, completely replace <Subject 3>, <Subject 5>, <Subject 7>, <Subject 9>, <Subject 11>, and <Subject 13> with <Subject 4>, <Subject 6>, <Subject 8>, <Subject 10>, <Subject 12>, and <Subject 14>, whose identities and complete visual appearances are defined by <Picture 2>, <Picture 3>, <Picture 4>, <Picture 5>, <Picture 6>, and <Picture 7> respectively.
Preserve the requested performance and scene structure from <Video 1>. Do not preserve, blend, or reintroduce the visual identity, face, hairstyle, body appearance, or outfit colors of any source model in <Video 1>.

retention_analysis:
<Video 1>: fully_preserved for body movement, poses, action timing, camera trajectory, framing, background, lighting, spatial relationships, group formation, and interaction with the environment. The identity appearance of every source model is excluded.
<Subject 1>: identity_not_preserved. Completely remove the red-outfit protagonist's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 1>: fully_referenced as the identity and complete visual-appearance source for <Subject 2>.
<Subject 2>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 1> consistent from every visible angle and in every frame.
<Subject 3>: identity_not_preserved. Completely remove the first white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 2>: fully_referenced as the identity and complete visual-appearance source for <Subject 4>.
<Subject 4>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 2> consistent from every visible angle and in every frame.
<Subject 5>: identity_not_preserved. Completely remove the second white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 3>: fully_referenced as the identity and complete visual-appearance source for <Subject 6>.
<Subject 6>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 3> consistent from every visible angle and in every frame.
<Subject 7>: identity_not_preserved. Completely remove the third white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 4>: fully_referenced as the identity and complete visual-appearance source for <Subject 8>.
<Subject 8>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 4> consistent from every visible angle and in every frame.
<Subject 9>: identity_not_preserved. Completely remove the fourth white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 5>: fully_referenced as the identity and complete visual-appearance source for <Subject 10>.
<Subject 10>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 5> consistent from every visible angle and in every frame.
<Subject 11>: identity_not_preserved. Completely remove the fifth white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 6>: fully_referenced as the identity and complete visual-appearance source for <Subject 12>.
<Subject 12>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 6> consistent from every visible angle and in every frame.
<Subject 13>: identity_not_preserved. Completely remove the sixth white-outfit model's face, hairstyle, body appearance, clothing, and visual identity.
<Picture 7>: fully_referenced as the identity and complete visual-appearance source for <Subject 14>.
<Subject 14>: identity_fully_preserved. Keep the identity and complete appearance from <Picture 7> consistent from every visible angle and in every frame.

reference_interpretation:
<Picture 1> through <Picture 7> are seven character turnaround sheets. Each sheet shows the same single person from three angles: a front view, a side view, and a back view. The front, side, and back figures on one sheet are one and the same character, not three different people, and the seven sheets depict seven different characters, exactly one character per sheet. Use the front view as the primary identity anchor for face and costume, the side view for profile, hair volume, and body silhouette, and the back view only for the back of the hairstyle and the back of the costume. Do not reproduce the sheet layout, panel dividers, text labels, grid lines, blank sheet background, or duplicate copies of a character.

detailed_description:
For the target video, at 0.00 seconds, <Picture 1> is fully referenced as the identity and complete visual appearance of <Subject 2>, and <Picture 2> through <Picture 7> are fully referenced as the identity and complete visual appearance of <Subject 4>, <Subject 6>, <Subject 8>, <Subject 10>, <Subject 12>, and <Subject 14> respectively.

[Shot 1] The target video begins with the same shot composition as <Video 1>. Exactly seven characters appear in the target video, occupying the same positions and the same left-to-right group order as the seven models in <Video 1>: <Subject 2> from <Picture 1> stands where the red-outfit protagonist <Subject 1> stands as the leading figure, and <Subject 4>, <Subject 6>, <Subject 8>, <Subject 10>, <Subject 12>, and <Subject 14> from <Picture 2> through <Picture 7> stand where the white-outfit supporting models <Subject 3>, <Subject 5>, <Subject 7>, <Subject 9>, <Subject 11>, and <Subject 13> stand, in exactly the same order. Every target character follows the same body movement, poses, physical actions, reactions, and action timing shown in <Video 1>. Preserve the camera movement, framing, environment, lighting, spatial relationships, group formation, and scene composition from <Video 1>. Only the designated characters' identities and appearances are replaced.

Throughout the target video, each target character remains one consistent single person whose identity and complete appearance come from its own sheet: <Subject 2> from <Picture 1>, <Subject 4> from <Picture 2>, <Subject 6> from <Picture 3>, <Subject 8> from <Picture 4>, <Subject 10> from <Picture 5>, <Subject 12> from <Picture 6>, and <Subject 14> from <Picture 7>. Never merge, swap, average, or blend one sheet's character into another. Do not retain or blend any characteristic from <Subject 1>, <Subject 3>, <Subject 5>, <Subject 7>, <Subject 9>, <Subject 11>, or <Subject 13>, and do not keep the red outfit or the plain white outfits of the source models. Keep exactly seven people in frame at all times, with no additional person, no missing person, no duplicate character, and no extra copies of the turnaround sheets. Each turnaround sheet is one character shown from three angles; never render one sheet as three separate people. Avoid identity mixing, face morphing, duplicated limbs, unstable anatomy, temporal flicker, and sudden appearance changes.

overall_soundscape:
Generate natural synchronized ambience and physical sounds consistent with the visible actions, group movement, and environment of <Video 1>. Keep the group's footsteps, clothing movement, and room tone coherent with the visible timing. Do not copy or reference the source video's audio unless it is explicitly wired or requested.

non_diegetic_music:
N/A
```