Watermark Eraser

High-precision, 100% coverage watermark removal tool powered by a V3 detector and LaMa inpainting.

This tool automatically detects and erases watermarks from images with pixel-perfect precision. It can be used as a standalone Python project or as a Manus Skill for direct integration with Code Agents.

---

🚀 Usage with Code Agents (Manus Skill)

This is the recommended way to use the tool for automated, large-scale processing. A Code Agent can directly use this repository as a skill to process images.

How It Works

1. Provide the GitHub URL: Give the agent the URL to this repository: https://github.com/moose-lab/watermark-eraser.git
2. Assign the Task: Instruct the agent to process a directory of images.

Example Prompt for a Code Agent

"Use the skill from https://github.com/moose-lab/watermark-eraser.git to remove the watermarks from all images in the /path/to/my/images directory. Save the clean images to /path/to/my/output."

Agent Workflow

The agent will automatically perform the following steps:

1. Clone the Repository: git clone https://github.com/moose-lab/watermark-eraser.git
2. Install Dependencies: pip install -r requirements.txt
3. Execute Batch Processing: Run the batch_processor.py script with the specified input and output directories.

# Agent will run this command
python watermark-eraser/src/batch_processor.py \
  --input /path/to/my/images/ \
  --output /path/to/my/output/ \
  --batch

4. Deliver Results: The agent will provide a link to the output directory containing:
   • Clean, watermark-free images.
   • A detailed processing_report.json.
   • An output_files.txt list for easy integration.

---

本地部署 (Local Deployment for Agents)

For advanced users who want to pre-install the skill in their local environment.

1. One-Step Installation

Run this command to clone the repository and install all dependencies:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/moose-lab/watermark-eraser/main/install.sh)"

This will:

• Clone the repository to /home/ubuntu/skills/watermark-eraser.
• Install all necessary Python packages.
• Download the LaMa model.

2. Instruct Your Agent

Once installed, you can instruct your agent to use the local skill:

"Use the local watermark-remover skill to process the images in /input/dir/ and save them to /output/dir/."

3. Manual Installation

For detailed steps, see the Local Deployment Guide.

---

✨ Key Features

• 100% Coverage V3 Detector: A custom-built, three-stage detector that guarantees every pixel of the target watermark is identified.
• Zero Residue: By providing a perfect mask to the inpainting model, it leaves no dark edges or blurry artifacts.
• LaMa Inpainting: Utilizes the state-of-the-art LaMa (Large Mask Inpainting) model for seamless, high-quality background reconstruction.
• Batch Processing: Process entire directories of images with a single command.
• Detailed Reporting: Generates a JSON report with statistics for each image, including mask coverage and processing time.
• GPU Acceleration: Supports CUDA and MPS for significantly faster processing.

🛠️ Standalone Usage

For local development or integration into non-agent workflows.

1. Installation

# Clone the repository
git clone https://github.com/moose-lab/watermark-eraser.git
cd watermark-eraser

# Install dependencies
pip install -r requirements.txt

2. Basic Usage

Process a single image:

python src/batch_processor.py \
  --input /path/to/your/image.jpg \
  --output /path/to/output/result.jpg

Process a directory of images:

python src/batch_processor.py \
  --input /path/to/your/images/ \
  --output /path/to/output/results/ \
  --batch

🔧 Advanced Options

GPU Acceleration

Use the --device flag to specify a processing device (cuda, mps, or cpu).

# Use NVIDIA GPU
python src/batch_processor.py --batch --device cuda ...

Quality Assurance

Save the detection mask and a visualization image for quality checks.

python src/batch_processor.py \
  --input ./images/ \
  --output ./qa_results/ \
  --batch \
  --save-masks \
  --visualize

📚 Documentation

• Local Deployment Guide: Step-by-step instructions for setting up the skill locally.
• Architecture: A detailed look at the system components and data flow.
• Development History: The story of how the V3 detector was built.
• Examples: Basic and advanced usage scripts.

📊 Performance

Based on a test set of 72 images with consistent watermarks:

Metric	Result
Success Rate	100% (72/72)
Mask Coverage	100.00%
False Positives	0%
Residue/Artifacts	0%
Avg. Time (CPU)	5.8 sec/image

🤝 Contributing

Contributions are welcome! Please feel free to submit a pull request or open an issue.

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.