watermark-removal

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Universal Watermark Removal

通用水印移除工具

Remove watermarks from ANY image source using intelligent detection and proven methods. Smart routing: automatically detects Google SynthID and uses the proven aggressive method, falls back to ML inpainting for unknown watermark types.

借助智能检测与成熟的处理方式，移除任意图片来源的水印。智能路由机制：自动检测Google SynthID水印并采用经验证的aggressive处理方式，针对未知类型水印则默认使用ML修复方式。

Quick Start

快速开始

Single Image - Smart Auto-Detection (Recommended)

单张图片 - 智能自动检测（推荐）

bash

undefined

bash

undefined

Smart detection: Google SynthID → aggressive method (proven), Unknown → inpaint (ML)

智能检测：Google SynthID → aggressive方式（经验证有效），未知水印 → inpaint（ML修复）

python .claude/skills/watermark-removal/scripts/remove-watermark.py
input.png
output.png

undefined

python .claude/skills/watermark-removal/scripts/remove-watermark.py
input.png
output.png

undefined

Single Image - Preserve Dimensions (Force ML)

单张图片 - 保留尺寸（强制使用ML修复）

bash

undefined

bash

undefined

Force ML inpainting even for Google SynthID (preserves exact dimensions)

即使是Google SynthID水印也强制使用ML修复（保留精确尺寸）

python .claude/skills/watermark-removal/scripts/remove-watermark.py
input.png
output.png
--method inpaint

undefined

python .claude/skills/watermark-removal/scripts/remove-watermark.py
input.png
output.png
--method inpaint

undefined

Batch Processing (Get User Approval First!)

批量处理（需先获得用户许可！）

⚠️ IMPORTANT: Always ask user before batch processing, especially with crop/aggressive methods that alter dimensions.

bash

undefined

⚠️ 重要提示： 批量处理前务必询问用户，尤其是使用会改变尺寸的crop/aggressive方式时。

bash

undefined

Recommended: Preserves dimensions

推荐：保留尺寸

python .claude/skills/watermark-removal/scripts/batch-process.py
/path/to/input-dir
/path/to/output-dir
--method inpaint

Alternative: Fast but crops 120px (requires user approval)

备选：快速处理但会裁剪120像素（需用户许可）

python .claude/skills/watermark-removal/scripts/batch-process.py
/path/to/input-dir
/path/to/output-dir
--method aggressive

undefined

python .claude/skills/watermark-removal/scripts/batch-process.py
/path/to/input-dir
/path/to/output-dir
--method aggressive

undefined

Smart Detection System ⭐ NEW

智能检测系统 ⭐ 新增功能

The skill automatically detects Google SynthID watermarks and routes to the optimal removal method:

该工具可自动检测Google SynthID水印，并选择最优移除方式：

Google SynthID Detection

Google SynthID检测

Characteristics analyzed:

RGBA mode (PNG format with alpha channel)
Large dimensions (>1500px width and height)
Typical Google AI aspect ratios (1.83, 1.0, 1.5, 1.78 with 10% tolerance)
Automatic corner detection for watermark location

Smart routing logic:

If Google SynthID detected → Uses
```
aggressive
```
method (proven to work perfectly)
- Crops 120px from detected corner
- Removes alpha channel watermarking
- Paints over any remnants
- Works 100% reliably on Google AI images
If unknown/other watermark → Uses
```
inpaint
```
method (ML-based)
- Preserves exact dimensions
- Uses OpenCV Navier-Stokes algorithm
- Works on any watermark type

分析特征：

RGBA模式（带透明通道的PNG格式）
大尺寸（宽高均>1500像素）
Google AI典型宽高比（1.83、1.0、1.5、1.78，误差容忍度10%）
自动检测水印所在角落位置

智能路由逻辑：

检测到Google SynthID水印 → 使用
```
aggressive
```
方式（经验证可完美移除）
- 从检测到的角落裁剪120像素
- 移除透明通道水印
- 修复残留痕迹
- 对Google AI图片的处理成功率100%
未知/其他类型水印 → 使用
```
inpaint
```
方式（基于ML）
- 保留精确尺寸
- 使用OpenCV Navier-Stokes算法
- 支持任意类型水印

Override Default Behavior

覆盖默认行为

bash

undefined

bash

undefined

Force ML inpainting even for Google SynthID (preserves dimensions)

即使是Google SynthID水印也强制使用ML修复（保留尺寸）

python scripts/remove-watermark.py input.png output.png --method inpaint

Force aggressive method for non-Google watermarks

对非Google水印强制使用aggressive方式

python scripts/remove-watermark.py input.png output.png --method aggressive

Disable auto-detection (assume bottom-right corner)

禁用自动检测（默认假设水印在右下角）

python scripts/remove-watermark.py input.png output.png --no-detect

undefined

python scripts/remove-watermark.py input.png output.png --no-detect

undefined

Methods

处理方式

Inpaint Method (Best Quality) ⭐ NEW

Inpaint方式（画质最佳）⭐ 新增功能

What it does: ML-based inpainting with automatic watermark detection

Features:

Automatically detects watermark location (any corner)
Uses OpenCV's Navier-Stokes inpainting algorithm
Intelligently fills watermark area with surrounding patterns
Works on ANY watermark type (not just Google SynthID)

Pros:

Highest quality results
Preserves exact dimensions
Works on watermarks in any corner
Handles complex backgrounds intelligently
Universal - works on all watermark types

Cons:

Requires OpenCV installation (
```
pip install opencv-python
```
)
Slightly slower than crop method
May need parameter tuning for very large watermarks

Use when:

You need the best possible quality
Watermark is on complex/detailed background
Preserving exact dimensions is critical
Working with non-Google watermarks

工作原理： 基于ML的修复技术，搭配自动水印检测

特性：

自动检测水印位置（任意角落）
使用OpenCV的Navier-Stokes修复算法
智能利用周边图案填充水印区域
支持任意类型水印（不仅限于Google SynthID）

优势：

画质最佳
保留精确尺寸
支持任意角落的水印
智能处理复杂背景
通用性强，支持所有水印类型

劣势：

需要安装OpenCV（
```
pip install opencv-python
```
）
比crop方式稍慢
超大水印可能需要调整参数

适用场景：

追求最佳画质
水印位于复杂/细节丰富的背景上
保留精确尺寸至关重要
处理非Google水印

Aggressive Method (Fast & Reliable)

Aggressive方式（快速可靠）

What it does: Auto-detects corner, crops 120px, removes alpha channel, paints remnants

Pros:

Fast and reliable
Automatic detection of watermark corner
Handles RGBA images properly
Good for batch processing

Cons:

Reduces image size by 120px
May crop content near edges

Use when:

Processing many images quickly (default for batch)
Size reduction is acceptable
Google SynthID watermarks

工作原理： 自动检测角落，裁剪120像素，移除透明通道，修复残留痕迹

优势：

快速可靠
自动检测水印所在角落
正确处理RGBA图片
适合批量处理

劣势：

图片尺寸缩小120像素
可能裁剪边缘附近的内容

适用场景：

快速处理大量图片（批量处理默认方式）
尺寸缩小可接受
处理Google SynthID水印

Crop Method

Crop方式

What it does: Auto-detects and crops watermark corner

Pros:

Fastest method
Automatic detection
Minimal processing

Cons:

May leave watermark remnants
Doesn't handle alpha channel watermarking

Use when:

Speed is top priority
Quick preview needed

工作原理： 自动检测并裁剪水印所在角落

优势：

速度最快
自动检测
处理步骤最少

劣势：

可能残留水印痕迹
无法处理透明通道水印

适用场景：

优先考虑速度
需要快速预览

Paint Method

Paint方式

What it does: Paints over watermark without cropping (no auto-detection)

Pros:

Preserves dimensions
Simple approach

Cons:

Assumes bottom-right corner only
May leave visible artifacts
Less reliable than inpaint

Use when:

Simple watermarks on solid backgrounds
Legacy compatibility

工作原理： 在不裁剪的情况下覆盖水印（无自动检测）

优势：

保留尺寸
操作简单

劣势：

仅默认水印在右下角
可能留下可见痕迹
可靠性低于inpaint方式

适用场景：

纯色背景上的简单水印
兼容旧版需求

Important Guidelines

重要指南

Dimension Preservation Priority

尺寸保留优先级

BALANCE DOMAIN KNOWLEDGE WITH DIMENSION PRESERVATION

Smart default behavior:
- Google SynthID detected →
```
aggressive
```
  method (proven perfect, crops 120px)
- Unknown watermark →
```
inpaint
```
  method (preserves dimensions)
User override available:
- Force dimension preservation with
```
--method inpaint
```
  flag
- Force cropping with
```
--method aggressive
```
  flag
User approval required for batch cropping:
- If processing multiple Google SynthID images with aggressive method
- Explain that method will reduce image size by 120px
- Get explicit confirmation
- Show before/after dimensions

平衡领域知识与尺寸保留需求

智能默认行为：
- 检测到Google SynthID水印 → 使用
```
aggressive
```
  方式（经验证完美，裁剪120像素）
- 未知类型水印 → 使用
```
inpaint
```
  方式（保留尺寸）
支持用户覆盖：
- 使用
```
--method inpaint
```
  参数强制保留尺寸
- 使用
```
--method aggressive
```
  参数强制裁剪
批量裁剪需用户许可：
- 若使用aggressive方式批量处理Google SynthID图片
- 需说明该方式会将图片尺寸缩小120像素
- 获取明确确认
- 展示处理前后的尺寸

Batch Processing Protocol

批量处理流程

NEVER start batch processing without user confirmation:

Show what will happen:
- Number of images to process
- Method to be used
- Whether dimensions will be preserved or altered
Get explicit approval:
- "I will process X images using [method]. This [will/will not] alter dimensions. Proceed?"
Prefer non-destructive:
- Default to
```
inpaint
```
  for batch processing
- Only use
```
aggressive
```
  if user specifically requests speed over quality

未经用户确认绝不要启动批量处理：

告知处理内容：
- 待处理图片数量
- 将要使用的处理方式
- 是否会改变尺寸
获取明确许可：
- “我将使用[方式]处理X张图片，该方式[会/不会]改变尺寸。是否继续？”
优先选择非破坏性方式：
- 批量处理默认使用
```
inpaint
```
  方式
- 仅当用户明确要求速度优先于画质时才使用
```
aggressive
```
  方式

Workflow

工作流程

1. Identify Watermarked Images

1. 识别带水印的图片

Common watermark types:

Google SynthID: Small star/sparkle icon in corner
Stock photos: Logo or text overlay
AI services: Corner badges (Midjourney, DALL-E)
Camera watermarks: Date/time stamps

常见水印类型：

Google SynthID： 角落的小星星/闪光图标
图库图片： 商标或文字叠加层
AI服务： 角落标识（Midjourney、DALL-E）
相机水印： 日期/时间戳

2. Choose Method

2. 选择处理方式

Smart Default Workflow (NEW):

Run script without --method flag - Smart detection automatically routes to best method
Google SynthID detected → Uses
```
aggressive
```
method (proven perfect)
Unknown watermark → Uses
```
inpaint
```
method (ML-based, preserves dimensions)
Override with --method flag if needed

Method Selection Guide:

Smart auto (recommended): No flag (detects Google SynthID → aggressive, else → inpaint)
Force dimension preservation:
```
--method inpaint
```
(ML-based, works on any watermark)
Force Google method:
```
--method aggressive
```
(crops 120px, perfect for SynthID)
Maximum speed:
```
--method crop
```
(fastest, crops but may leave remnants)
Legacy:
```
--method paint
```
(basic, preserves dimensions but less reliable)

智能默认流程（新增）：

不带--method参数运行脚本 - 智能检测自动选择最佳方式
检测到Google SynthID水印 → 使用
```
aggressive
```
方式（经验证完美）
未知类型水印 → 使用
```
inpaint
```
方式（基于ML，保留尺寸）
必要时使用--method参数覆盖默认设置

处理方式选择指南：

智能自动（推荐）： 不带参数（检测到Google SynthID→aggressive，其他→inpaint）
强制保留尺寸：
```
--method inpaint
```
（基于ML，支持任意水印）
强制使用Google专属方式：
```
--method aggressive
```
（裁剪120像素，完美处理SynthID）
最快速度：
```
--method crop
```
（速度最快，裁剪但可能残留痕迹）
旧版兼容：
```
--method paint
```
（基础方式，保留尺寸但可靠性较低）

3. Process Images

3. 处理图片

Smart auto-detection (recommended):

bash

python scripts/remove-watermark.py input.png output.png

Force dimension preservation:

bash

python scripts/remove-watermark.py input.png output.png --method inpaint

Fast batch processing:

bash

python scripts/batch-process.py ./input ./output --method aggressive

Disable auto-detection (force bottom-right):

bash

python scripts/remove-watermark.py input.png output.png --method inpaint --no-detect

智能自动检测（推荐）：

bash

python scripts/remove-watermark.py input.png output.png

强制保留尺寸：

bash

python scripts/remove-watermark.py input.png output.png --method inpaint

快速批量处理：

bash

python scripts/batch-process.py ./input ./output --method aggressive

禁用自动检测（强制假设水印在右下角）：

bash

python scripts/remove-watermark.py input.png output.png --method inpaint --no-detect

4. Verify Results

4. 验证结果

Check output images for clean corners
Verify no important content was cropped
Confirm watermark fully removed

检查输出图片的角落是否干净
确认边缘附近的重要内容未被裁剪
确认水印已完全移除

Command Reference

命令参考

remove-watermark.py

bash

python scripts/remove-watermark.py INPUT OUTPUT [OPTIONS]

Arguments:
  INPUT                 Input image path
  OUTPUT                Output image path

Options:
  --method {crop|inpaint}   Removal method (default: crop)
  --size INT               Watermark size in pixels (default: 60)

bash

python scripts/remove-watermark.py INPUT OUTPUT [OPTIONS]

参数：
  INPUT                 输入图片路径
  OUTPUT                输出图片路径

选项：
  --method {crop|inpaint}   移除方式（默认：crop）
  --size INT               水印尺寸（像素，默认：60）

batch-process.py

bash

python scripts/batch-process.py INPUT_DIR OUTPUT_DIR [OPTIONS]

Arguments:
  INPUT_DIR             Directory with images to process
  OUTPUT_DIR            Directory for cleaned images

Options:
  --method {crop|inpaint}   Removal method (default: crop)
  --pattern PATTERN        File pattern to match (default: *.png)
  --size INT               Watermark size in pixels (default: 60)

bash

python scripts/batch-process.py INPUT_DIR OUTPUT_DIR [OPTIONS]

参数：
  INPUT_DIR             待处理图片目录
  OUTPUT_DIR            清理后图片的保存目录

选项：
  --method {crop|inpaint}   移除方式（默认：crop）
  --pattern PATTERN        匹配的文件格式（默认：*.png）
  --size INT               水印尺寸（像素，默认：60）

Examples

示例

Example 1: Website Images

示例1：网站图片处理

Scenario: User has 3 Google AI images for website, watermarks need removal

bash

undefined

场景： 用户有3张用于网站的Google AI图片，需要移除水印

bash

undefined

Batch process all PNG images

批量处理所有PNG图片

python scripts/batch-process.py
./public/images
./public/images-clean
--method crop
--pattern "*.png"


**Output:**

Found 3 images to process Input: ./public/images Output: ./public/images-clean Method: crop

[1/3] Processing: ad-design-bedroom.png ✅ Saved to: ad-design-bedroom.png [2/3] Processing: ad-design-kitchen.png ✅ Saved to: ad-design-kitchen.png [3/3] Processing: ad-designs-bathroom.png ✅ Saved to: ad-designs-bathroom.png

✅ Successfully processed: 3/3

undefined

python scripts/batch-process.py
./public/images
./public/images-clean
--method crop
--pattern "*.png"


**输出：**

找到3张待处理图片输入目录: ./public/images 输出目录: ./public/images-clean 处理方式: crop

[1/3] 处理中: ad-design-bedroom.png ✅ 保存至: ad-design-bedroom.png [2/3] 处理中: ad-design-kitchen.png ✅ 保存至: ad-design-kitchen.png [3/3] 处理中: ad-designs-bathroom.png ✅ 保存至: ad-designs-bathroom.png

✅ 处理完成: 3/3

undefined

Example 2: Preserve Exact Dimensions

示例2：保留精确尺寸

Scenario: Client needs exact 1920x1080 image, can't crop

bash

python scripts/remove-watermark.py \
  hero-image.png \
  hero-image-clean.png \
  --method inpaint

场景： 客户需要精确的1920x1080图片，不能裁剪

bash

python scripts/remove-watermark.py \
  hero-image.png \
  hero-image-clean.png \
  --method inpaint

Example 3: Larger Watermark

示例3：超大水印处理

Scenario: Watermark is bigger than usual (70px)

bash

python scripts/batch-process.py \
  ./images \
  ./images-clean \
  --method crop \
  --size 70

场景： 水印尺寸比常规大（70像素）

bash

python scripts/batch-process.py \
  ./images \
  ./images-clean \
  --method crop \
  --size 70

Technical Details

技术细节

See references/synthid-watermark.md for:

SynthID watermark specifications
Method comparison details
Edge cases and considerations
Legal/ethical guidelines

查看references/synthid-watermark.md获取：

SynthID水印规格
处理方式对比细节
边缘情况与注意事项
法律/伦理指南

Dependencies

依赖项

Required:

Python 3.7+
Pillow (PIL):
```
pip install Pillow
```
NumPy:
```
pip install numpy
```

Optional (for ML inpainting):

OpenCV:
```
pip install opencv-python
```

Install all dependencies:

bash

pip install Pillow numpy opencv-python

Note: The

inpaint

method requires OpenCV. Other methods work without it.

必需：

Python 3.7+
Pillow (PIL):
```
pip install Pillow
```
NumPy:
```
pip install numpy
```

可选（用于ML修复）：

OpenCV:
```
pip install opencv-python
```

安装所有依赖项：

bash

pip install Pillow numpy opencv-python

注意：

inpaint

方式需要OpenCV。其他方式无需安装。

Tips

技巧

Performance:

Crop method is 5-10x faster than inpaint
For 100+ images, use crop method

Quality:

Save with quality=95 to minimize compression
PNG format preserves quality better than JPEG

Backup:

Always keep original watermarked images
Process copies, not originals

Testing:

Test on one image before batch processing
Verify watermark size with
```
--size
```
flag if needed

性能：

Crop方式比inpaint方式快5-10倍
处理100张以上图片时，使用crop方式

画质：

保存时设置quality=95以最小化压缩损失
PNG格式比JPEG格式更能保留画质

备份：

始终保留带水印的原始图片
处理副本而非原图

测试：

批量处理前先测试单张图片
必要时使用
```
--size
```
参数验证水印尺寸