Watermark Remover
https://github.com/santifer/watermark-remover- Category
- AI Tools
- Rank
- No. 1638Tools index
Previous survey · No. 1646 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- santifer
- GitHub
- 25 stars
- Date
About
CLI tool to remove watermarks from images using YOLO detection plus LaMa inpainting in a single pipeline.
What it does
Give it an image, and it creates a cleaned copy by finding regions to replace, building a rectangular mask, then reconstructing those pixels. You can bypass automatic detection and target a chosen corner, or select a faster reconstruction method for simpler images.
Why it's ranked here
The command offers useful controls and two reconstruction choices, but its automatic masking logic is risky. The bundled detector uses a general YOLO model and accepts every returned object box, without checking whether the detected class represents a watermark. That can erase legitimate subjects.
What's good
The operator can tune detection confidence, mask padding, fallback corner, and corner dimensions. Detection failures do not necessarily stop processing because the tool can target a configurable corner. LaMa favors reconstruction quality, while OpenCV provides a faster alternative. Verbose output reports detections and bounding boxes.
Tradeoffs
Automatic fallback can alter an unmarked corner when detection finds nothing. More seriously, every YOLO detection becomes part of the removal mask, so ordinary detected objects may be reconstructed. LaMa runs on CPU, needs roughly 2 GB of memory, and downloads a model of about 200 MB on first use.
How to use it well
It best suits engineers processing individual images where they can inspect every result. Prefer forced-corner mode when watermark placement is predictable, and tune the mask dimensions conservatively. Use OpenCV for simple patches and LaMa when reconstruction matters more. It does not provide a built-in batch workflow, so larger jobs require external scripting.
Technical notes+
setup.py requires Python 3.10+, declares the watermark-remover console entry point, and installs Torch, TorchVision, Ultralytics, Pillow, OpenCV, NumPy, Click, and simple-lama-inpainting. watermark_remover/cli.py accepts one input path, converts it to RGB, builds a mask, inpaints, and saves a derived or explicit output path. watermark_remover/detector.py loads yolov8n.pt unless given an existing model path, returns every model box, and draws padded rectangular masks without class filtering. watermark_remover/inpainter.py clears CUDA_VISIBLE_DEVICES, manually downloads big-lama.pt through Torch Hub, loads it as TorchScript on CPU, and alternatively runs OpenCV Navier-Stokes inpainting. watermark_remover/__init__.py exports WatermarkDetector and WatermarkInpainter for Python use.
Observed
- License
- MIT
- Primary language
- Python
- Python requirement
- Python 3.10 or newer
- Installation surface
- Editable pip installation from a cloned repository
- Interfaces
- Command-line entry point, module execution, and importable detector and inpainter classes
- Core dependencies
- PyTorch, TorchVision, Ultralytics, Pillow, OpenCV, NumPy, Click, and simple-lama-inpainting
- Test structure
- No test directory appears in the provided repository tree
Read from README.md, setup.py, requirements.txt, watermark_remover/cli.py, watermark_remover/__init__.py, watermark_remover/__main__.py, watermark_remover/detector.py, watermark_remover/inpainter.py, LICENSE, README.es.md.
What it can do
Remove watermarks from images
Image files with watermarks → Clean images without watermarks
Detect watermarks in images
Image files → Watermark location coordinates
Inpaint detected watermark areas
Images with identified watermark regions → Images with watermark areas filled with contextually appropriate content
Process images through command line interface
Command line arguments and image file paths → Processed images saved to specified location
Run automated watermark removal pipeline
Batch of watermarked images → Batch of cleaned images
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.