refactor: make Liang skin image-only

This commit is contained in:
Leon
2026-08-18 19:03:18 +08:00
parent cfaa104fcb
commit 4060b9fb84
11 changed files with 22 additions and 144 deletions

Binary file not shown.

Binary file not shown.

View File

@@ -1,4 +1,4 @@
# 滑动变祖 · 人像与视频生成交接
# 滑动变祖 · 人像生成交接
## 结论
@@ -6,18 +6,17 @@
- 已生成:**24 / 24**
- 分辨率与色彩:**1024 × 1024,RGB**
- 生成图目录:`packages/dsh-client-liang-intensity-skin/assets/portrait-source-v2/`
- 原视频关键帧目录:`packages/dsh-client-liang-intensity-skin/assets/portrait-source-v2/references/`
- 原视频:`packages/dsh-client-liang-intensity-skin/assets/liang-evolution.webm`
- 阶段参考图目录:`packages/dsh-client-liang-intensity-skin/assets/portrait-source-v2/references/`
`stage-low.png`、`stage-mid.png`、`stage-max.png` 是早期别名。后续生图和生视频只使用表格里的 24 个规范文件名。
`stage-low.png`、`stage-mid.png`、`stage-max.png` 是早期别名。后续只使用表格里的 24 个规范文件名。
## 公共提示词
下面这段作为每张图的公共提示词。把大括号变量替换为表格中对应内容;阶段服装、表情和装饰必须以该阶段原视频关键帧为准。
下面这段作为每张图的公共提示词。把大括号变量替换为表格中对应内容;阶段服装、表情和装饰必须以该阶段参考图为准。
```text
Use case: identity-preserve
Asset type: square key portrait for a continuous image-to-video character evolution sequence
Asset type: square key portrait for a continuous portrait evolution sequence
Primary request:
Using {REFERENCE_FRAME} as the authoritative stage reference, create the level {LEVEL} key portrait of the same Chinese adult man. Preserve his identity exactly: facial structure, glasses, hairstyle, skin tone, age, eye spacing, nose, mouth and overall proportions. Preserve the outfit, expression, posture and stage-specific accessories shown in the authoritative reference. Do not borrow clothing or accessories from a later stage.
@@ -65,34 +64,3 @@ No close-up crop. No cropped head, shoulders, arms or torso. No side pose. No ex
| 24 | 30 | 8.000s | 已生成 | `stage-30.png` | `references/stage-30-8.000s.png` | 前一张 `level-29.png` |
所有表格路径均相对于本文件所在目录。
## 生视频输入顺序
```text
stage-00.png
level-01.png
level-03.png
level-04.png
stage-06.png
level-07.png
level-09.png
level-10.png
stage-12.png
level-13.png
level-14.png
bridge-15.png
level-16.png
level-17.png
stage-18.png
level-19.png
level-21.png
level-22.png
stage-24.png
level-25.png
bridge-27.png
level-28.png
level-29.png
stage-30.png
```
视频 agent 应锁定镜头、构图和人物位置,只在相邻关键图之间做外观演化;不要加入推拉、横移、眨眼、转头、嘴部动作或随机肢体运动。输出仍按当前 UI 使用的 1:1 方形画幅制作,左侧留白和人物右移由 UI 的裁切、定位与渐变遮罩完成。