WatermarkRemover-AI: Open-Source Desktop Watermark Removal with Florence-2 + LaMA

Background

Most watermark removal tools are online services — you upload images to a server, wait for cloud processing, and download the results. That's fine for casual use, but when images contain sensitive or confidential content, sending them to a third-party server isn't ideal.

Microsoft's Florence-2 is a vision foundation model that excels at object detection and image understanding. The community has fine-tuned it specifically for watermark detection. LaMA (Large Mask Inpainting), on the other hand, is a well-established image inpainting method — it reconstructs the area underneath a watermark by analyzing surrounding pixels. Combining these two models gives us WatermarkRemover-AI, a recently open-sourced desktop application on GitHub.

Created by developer D-Ogi, this tool is completely free, MIT-licensed, supports GPU acceleration, and — most importantly — provides a graphical user interface so you don't need to write any code.

Core Architecture

The processing pipeline is straightforward:

1. Detection phase: Florence-2 scans the image or video frame to precisely locate watermark boundaries
2. Mask generation: Detected watermark regions are converted into a mask (adjustable in size)
3. Inpainting phase: LaMA intelligently fills the masked area based on surrounding pixels
4. Output phase: The cleaned file is saved; for videos, the original audio is preserved

The two-phase design has a key advantage — detection and inpainting can be optimized independently. You could swap Florence-2 for another detection model or LaMA for a stronger inpainting model without affecting the other component.

Three Ways to Use It

1. GUI Mode: Out of the Box

This is the recommended approach. After installation, run run.bat (Windows) or ./run.sh (Linux/macOS). A desktop window appears:

• Select language (English, Chinese, Japanese, French, and more)
• Choose a theme
• Pick single-file or batch mode
• Set input/output paths
• Click "Start Processing"

Settings are auto-saved for the next session.

2. CLI Mode: For Automation

Perfect for scripting and CI/CD pipelines:

# Basic usage
python remwm.py input.mp4 output.mp4

# With options
python remwm.py input.mp4 output.mp4 --overwrite --fade-in 2 --fade-out 3

# Preview detected watermarks (no actual processing)
python remwm.py input.jpg output.jpg --preview

3. Batch Mode

Process entire folders at once:

python remwm.py /path/to/input/folder /path/to/output/folder --overwrite

Video Watermark Removal: Frame Detection + Fade Handling

Video watermark removal is more complex than images — watermarks may shift slightly between frames, and often have fade-in/fade-out effects. WatermarkRemover-AI handles these challenges specifically:

• Two-pass detection: Scans every N frames for watermarks, balancing speed and accuracy
• Fade in/out handling: The --fade-in and --fade-out parameters extend the mask to cover transition frames
• Audio preservation: Original audio track is retained after processing (requires FFmpeg)

Supported video formats include MP4, AVI, MOV, MKV, FLV, WMV, and WEBM.

Installation

Windows

The setup.bat script downloads a portable Python — no system Python required. After setup, double-click run.bat to launch.

Linux / macOS

Requires Python 3.10+:

cd WatermarkRemover-AI
pip install -r requirements.txt

Then run ./run.sh to start.

FFmpeg (Required for Video)

Video processing requires FFmpeg:

• Linux: sudo apt install ffmpeg
• macOS: brew install ffmpeg
• Windows: Download from ffmpeg.org and add to PATH

Performance

According to the project README, on a dual RTX 4090 machine, image processing reaches 1000+ images per minute. For everyday users, a single RTX 3060 or 4070 provides a smooth experience.

Without a dedicated GPU, CPU mode still works but is significantly slower — fine for occasional small batches.

Use Cases

• Privacy-sensitive watermark removal: All processing stays local — nothing is uploaded to any server
• AI-generated video cleanup: Removes watermarks from Sora, Sora 2, Runway, Kling, and other AI video platforms
• Personal media organization: Batch-clean watermarks from downloaded media collections
• Content creators: Remove corner badges from social media video downloads
• Dataset preprocessing: Clean training data of watermark artifacts (batch mode)

Comparison with Online Services

Dimension	WatermarkRemover-AI	Online Services
Processing	Local	Cloud upload
Privacy	Data never leaves your machine	Depends on provider's policy
Cost	Free (MIT License)	Free tier often limited
Video	Full support with audio	Partial support
GPU	CUDA acceleration	Handled server-side
Setup	Requires Python + dependencies	Browser only

CLI Options Reference

Option	Description	Default
--overwrite	Overwrite existing output files	Off
--transparent	Make watermark regions transparent (images only)	Off
--max-bbox-percent	Max detection box as % of image	10
--force-format	Force output format (PNG/WEBP/JPG/MP4/AVI)	Auto
--detection-prompt	Custom detection prompt	"watermark"
--detection-skip	Frame detection interval (1-10)	1
--fade-in	Extend mask backward (seconds)	0
--fade-out	Extend mask forward (seconds)	0
--preview	Preview only, no processing	Off

Tech Stack

• Detection model: Florence-2 (Microsoft)
• Inpainting model: LaMA (Large Mask Inpainting)
• GUI framework: PyWebview + Alpine.js
• Deep learning backend: PyTorch + CUDA

Florence-2's general vision capabilities allow WatermarkRemover-AI to recognize various watermark types — not just fixed patterns from specific platforms. LaMA handles large masked areas reliably, leaving minimal visible traces after watermark removal.

Summary

WatermarkRemover-AI is a noteworthy open-source project in the 2026 watermark removal landscape. By combining Microsoft's vision model with a proven inpainting approach, it delivers a cross-platform desktop application that prioritizes user privacy. For anyone who needs local, batch-capable watermark removal — especially for AI-generated video content — this tool is worth trying.

The only barrier is the Python dependency setup, but Windows users get a one-click installer, and Linux/macOS users need just a few commands.

Project page: https://github.com/D-Ogi/WatermarkRemover-AI