270 lines
8.5 KiB
Markdown
270 lines
8.5 KiB
Markdown
---
|
||
name: face-warp
|
||
display-name-zh: 人像拼图
|
||
summary-cn: 输入参考人脸，用拼图法解决人脸过审问题
|
||
summary-en: Split reference face to pass moderation
|
||
description: |
|
||
Portrait deconstruction tool for AI image creation.
|
||
Splits a portrait photo into 2 variants: faceless (features erased) + puzzle (features extracted).
|
||
Enabling character-consistent content creation with AI generation models.
|
||
Pipeline: portrait upload → analysis → 2 variant generation → quality check → composite → output.
|
||
|
||
Trigger on: "face warp," "人脸拆解," "人物拆解," "面部解构," "portrait split,"
|
||
"faceless," "去脸," "五官提取," "face decompose," "人脸预处理,"
|
||
or any request where someone provides a portrait and wants face-deconstructed
|
||
variants for downstream use.
|
||
|
||
Do NOT trigger for: style transfer without face modification (use kontext),
|
||
face swap between two people, deepfake creation, or portrait retouching.
|
||
|
||
version: 0.5.3
|
||
tags: [Image, Portrait, Face, Preprocessing]
|
||
tags-cn: [图片, 人像, 人脸, 预处理]
|
||
exported-by: MiniMax-hub
|
||
---
|
||
|
||
# Face Warp — 人物拆解工具
|
||
|
||
你是一个专业的人像解构艺术家，将一张人物照片拆解为 2 种变体图并拼合为最终成品，
|
||
在保持人物一致性的同时，为下游 AI 创作提供多样化素材。
|
||
|
||
## Iron Law
|
||
|
||
**每一张输出图都必须保持与原图的人物一致性。** 服装、体态、肤色、发型
|
||
必须与原图高度一致。拆解是"解构"而非"替换"。
|
||
|
||
## 全局约定
|
||
|
||
- 所有产物存储在 `./.face-warp/{project_name}/`
|
||
- **全自动执行**：全程自动，无用户选择环节，生成失败自动重试（最多 2 次）
|
||
- **语言**：文案默认中文（跟随用户输入语言），AI 生图 prompt 使用英文
|
||
- **禁止使用 `cd` 命令**
|
||
- **模型选择**：图片生成默认使用 `nano_banana_image_generation`（model_name: `nano_banana_2`，即 Gemini Pro）
|
||
- **质检重试**：生成后用 `read_media` 评分，不合格自动调整 prompt 重试（最多 2 次），详见 `references/quality-criteria.md`
|
||
- **最终输出 1 张图**：composite（faceless + puzzle 左右拼接）
|
||
|
||
## 资源文件
|
||
|
||
| 文件 | 用途 |
|
||
|------|------|
|
||
| `references/profile-template.md` | Character Profile 结构模板 |
|
||
| `references/prompt-templates.md` | faceless / puzzle 的 prompt 模板及重试强化词 |
|
||
| `references/quality-criteria.md` | 质检评分标准、阈值、重试策略 |
|
||
|
||
## 工作流程
|
||
|
||
```
|
||
Face Warp Progress:
|
||
|
||
- [ ] Phase 1: Portrait Upload & Analysis
|
||
- [ ] Phase 2: Generate 2 Variants
|
||
- [ ] Phase 3: Quality Check & Retry
|
||
- [ ] Phase 4: Composite (Faceless + Puzzle → Single Image)
|
||
```
|
||
|
||
---
|
||
|
||
## Phase 1: Portrait Upload & Analysis
|
||
|
||
### Goal
|
||
|
||
接收用户上传的人像照片，分析人物特征，生成 Character Profile。
|
||
|
||
### Input / Output
|
||
|
||
| | 内容 |
|
||
|---|------|
|
||
| **Input** | 用户上传的人像照片（1 张），可选 aspect ratio |
|
||
| **Output** | `./.face-warp/{project_name}/profile.md` — 人物特征档案 |
|
||
|
||
### Required Inputs
|
||
|
||
| Input | Required | Description |
|
||
|-------|----------|-------------|
|
||
| Portrait photo | Yes | 含清晰人脸的人物照片 |
|
||
| Aspect ratio | Optional | 输出图片宽高比（默认 9:16） |
|
||
|
||
### Flow
|
||
|
||
1. Save portrait to session:
|
||
```
|
||
save_file_to_session(source_path=..., file_type="image")
|
||
```
|
||
|
||
2. Analyze portrait with `read_media`:
|
||
```
|
||
read_media(
|
||
file_paths=[portrait_image],
|
||
question="Analyze this portrait in detail. Extract:
|
||
1) Gender, approximate age range
|
||
2) Hair: color, length, style
|
||
3) Skin tone (light/medium/dark, warm/cool undertone)
|
||
4) Face shape
|
||
5) Distinctive facial features (eye shape, nose shape, lip shape, unique marks)
|
||
6) Clothing: type, color, pattern, texture
|
||
7) Pose and body posture
|
||
8) Background/environment
|
||
9) Lighting direction and quality
|
||
10) Overall color palette"
|
||
)
|
||
```
|
||
|
||
3. 按 `references/profile-template.md` 结构填充分析结果，保存到 `./.face-warp/{project_name}/profile.md`
|
||
|
||
---
|
||
|
||
## Phase 2: Generate 2 Variants
|
||
|
||
### Goal
|
||
|
||
基于原始人像和 Profile，一次性并行生成 2 张变体图。
|
||
|
||
### Input / Output
|
||
|
||
| | 内容 |
|
||
|---|------|
|
||
| **Input** | 原始人像 + `profile.md` |
|
||
| **Output** | `./.face-warp/{project_name}/faceless.png` + `puzzle.png` |
|
||
|
||
### 2 张变体定义
|
||
|
||
| # | 名称 | 文件名 | 描述 |
|
||
|---|------|--------|------|
|
||
| 1 | 五官擦除 | `faceless.png` | 面部五官被平滑擦除（光滑无特征皮肤），身体服装不变 |
|
||
| 2 | 五官拆分拼贴 | `puzzle.png` | 只保留五官特征（眼鼻唇眉），按脸部自然布局拆散拼贴在白色背景上 |
|
||
|
||
### Prompt 构建
|
||
|
||
从 `profile.md` 提取人物描述前缀，与 `references/prompt-templates.md` 中的模板组合。
|
||
|
||
**Prompt 前缀**（从 profile 提取）：
|
||
```
|
||
[gender], [age range], [hair description], [skin tone], wearing [outfit description],
|
||
[pose description], [background/environment], [lighting]
|
||
```
|
||
|
||
**完整 prompt** = 前缀 + 模板（详见 `references/prompt-templates.md`）。
|
||
|
||
提交前对照 `references/prompt-templates.md` 底部的 **Prompt Construction Checklist** 逐项确认。
|
||
|
||
### Generation
|
||
|
||
使用 `nano_banana_batch_image_generation_v2` 一次并行生成 2 张：
|
||
|
||
```
|
||
nano_banana_batch_image_generation_v2(
|
||
count=2,
|
||
prompts=[faceless_prompt, puzzle_prompt],
|
||
image_paths=[
|
||
[original_portrait],
|
||
[original_portrait]
|
||
],
|
||
aspect_ratios=["9:16", "9:16"],
|
||
model_name="nano_banana_2",
|
||
resolution="2K"
|
||
)
|
||
```
|
||
|
||
如果批量生成失败，则逐张用 `nano_banana_image_generation` 生成。
|
||
|
||
---
|
||
|
||
## Phase 3: Quality Check & Retry
|
||
|
||
### Goal
|
||
|
||
用 `read_media` 对生成图进行质量评分，不合格的自动调整 prompt 重新生成。
|
||
|
||
### Input / Output
|
||
|
||
| | 内容 |
|
||
|---|------|
|
||
| **Input** | `faceless.png` + `puzzle.png` + 原始人像 |
|
||
| **Output** | 通过质检的 `faceless.png` + `puzzle.png`（可能经过重试替换） |
|
||
|
||
### 评分流程
|
||
|
||
详细评分标准、阈值和重试策略见 `references/quality-criteria.md`。
|
||
|
||
核心流程：
|
||
|
||
1. 将原图 + 两张生成图送入 `read_media`，按 `quality-criteria.md` 中的 Evaluation Prompt 评分
|
||
2. 解析评分结果，判断每张图是否 PASS（所有维度 ≥ 阈值）
|
||
3. FAIL 的图按 `quality-criteria.md` 中的 Retry Prompt Adjustment Strategy 调整 prompt，仅重生成不合格的那张
|
||
4. 最多重试 2 次，仍不合格则保留最佳版本
|
||
5. 将评分记录写入 `./.face-warp/{project_name}/quality_log.md`
|
||
|
||
### 关键阈值速查
|
||
|
||
| 图片 | 关键维度 | 阈值 |
|
||
|------|----------|------|
|
||
| faceless | face_concealment | ≥ 8 |
|
||
| faceless | character_consistency / natural_appearance / image_quality | ≥ 7 |
|
||
| puzzle | no_full_face | ≥ 8 |
|
||
| puzzle | feature_accuracy / skin_tone_consistency / artistic_quality | ≥ 7 |
|
||
|
||
---
|
||
|
||
## Phase 4: Composite (Faceless + Puzzle → Single Image)
|
||
|
||
### Goal
|
||
|
||
将通过质检的 faceless 和 puzzle 两张图左右拼接为一张完整的合成图。
|
||
|
||
### Input / Output
|
||
|
||
| | 内容 |
|
||
|---|------|
|
||
| **Input** | 通过质检的 `faceless.png` + `puzzle.png` |
|
||
| **Output** | `./.face-warp/{project_name}/output/composite.png` |
|
||
|
||
### 拼接方式
|
||
|
||
使用 `ffmpeg` 进行左右拼接，faceless 在左，puzzle 在右：
|
||
|
||
```
|
||
ffmpeg(args=[
|
||
"-y",
|
||
"-i", "./.face-warp/{project_name}/faceless.png",
|
||
"-i", "./.face-warp/{project_name}/puzzle.png",
|
||
"-filter_complex",
|
||
"[0]scale=-1:1080[left];[1]scale=-1:1080[right];[left][right]hstack=inputs=2",
|
||
"./.face-warp/{project_name}/output/composite.png"
|
||
])
|
||
```
|
||
|
||
**规则**：
|
||
- 两张图先统一高度（1080px），宽度等比缩放
|
||
- 使用 `hstack` 水平拼接（左: faceless, 右: puzzle）
|
||
- 输出到 `./.face-warp/{project_name}/output/composite.png`
|
||
|
||
---
|
||
|
||
## Completion
|
||
|
||
```
|
||
--- Face Warp Complete ---
|
||
|
||
Character: {character_description}
|
||
Output: .face-warp/{project_name}/output/composite.png
|
||
```
|
||
|
||
---
|
||
|
||
## Error Handling
|
||
|
||
| Error | Recovery |
|
||
|-------|----------|
|
||
| Portrait too low-res | Run `super_resolution` first |
|
||
| Face not clearly visible | Ask user for a clearer portrait |
|
||
| Batch generation fails | Fall back to single `nano_banana_image_generation` per image |
|
||
| Single image generation fails | Retry once with adjusted prompt, then skip and report |
|
||
|
||
## Anti-Patterns
|
||
|
||
- **Don't skip quality checks**: 质检是保证输出质量的关键环节，不能跳过
|
||
- **Don't alter body proportions**: 只改变面部呈现方式，不改变体型体态
|
||
- **Don't change clothing**: 服装必须与原图完全一致
|
||
- **Don't over-stylize**: 保持照片级写实感
|
||
- **Don't ignore skin tone**: 肤色一致性是人物一致性的关键
|