watermark-removal（水印去除）

Universal Watermark Removal

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.

Quick Start

Single Image - Smart Auto-Detection (Recommended)

# Smart detection: Google SynthID → aggressive method (proven), Unknown → inpaint (ML)
python .claude/skills/watermark-removal/scripts/remove-watermark.py \
  input.png \
  output.png

Single Image - Preserve Dimensions (Force ML)

# Force ML inpainting even for Google SynthID (preserves exact dimensions)
python .claude/skills/watermark-removal/scripts/remove-watermark.py \
  input.png \
  output.png \
  --method inpaint

Batch Processing (Get User Approval First!)

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

# Recommended: Preserves dimensions
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)
python .claude/skills/watermark-removal/scripts/batch-process.py \
  /path/to/input-dir \
  /path/to/output-dir \
  --method aggressive

Smart Detection System ⭐ NEW

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

Google SynthID Detection

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:

1. 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

1. If unknown/other watermark → Uses inpaint method (ML-based)

- Preserves exact dimensions - Uses OpenCV Navier-Stokes algorithm - Works on any watermark type

Override Default Behavior

# Force ML inpainting even for Google SynthID (preserves dimensions)
python scripts/remove-watermark.py input.png output.png --method inpaint

# Force aggressive method for non-Google watermarks
python scripts/remove-watermark.py input.png output.png --method aggressive

# Disable auto-detection (assume bottom-right corner)
python scripts/remove-watermark.py input.png output.png --no-detect

Methods

Inpaint Method (Best Quality) ⭐ NEW

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

Aggressive Method (Fast & Reliable)

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

Crop Method

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

Paint Method

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

Important Guidelines

Dimension Preservation Priority

BALANCE DOMAIN KNOWLEDGE WITH DIMENSION PRESERVATION

1. Smart default behavior:

- Google SynthID detected → aggressive method (proven perfect, crops 120px) - Unknown watermark → inpaint method (preserves dimensions)

1. User override available:

- Force dimension preservation with --method inpaint flag - Force cropping with --method aggressive flag

1. 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

Batch Processing Protocol

NEVER start batch processing without user confirmation:

1. Show what will happen:

- Number of images to process - Method to be used - Whether dimensions will be preserved or altered

1. Get explicit approval:

- "I will process X images using [method]. This [will/will not] alter dimensions. Proceed?"

1. Prefer non-destructive:

- Default to inpaint for batch processing - Only use aggressive if user specifically requests speed over quality

Workflow

1. Identify Watermarked Images

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

2. Choose Method

Smart Default Workflow (NEW):

1. Run script without --method flag - Smart detection automatically routes to best method
2. Google SynthID detected → Uses aggressive method (proven perfect)
3. Unknown watermark → Uses inpaint method (ML-based, preserves dimensions)
4. 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)

3. Process Images

Smart auto-detection (recommended):

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

Force dimension preservation:

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

Fast batch processing:

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

Disable auto-detection (force bottom-right):

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

4. Verify Results

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

Command Reference

remove-watermark.py

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)

batch-process.py

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)

Examples

Example 1: Website Images

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

# Batch process all PNG images
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

Example 2: Preserve Exact Dimensions

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

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

Example 3: Larger Watermark

Scenario: Watermark is bigger than usual (70px)

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

Technical Details

See references/synthid-watermark.md for:

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

Dependencies

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:

pip install Pillow numpy opencv-python

Note: The inpaint method requires OpenCV. Other methods work without it.

Tips

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