Video Watermark Removal CLI

Automatically detect and remove watermarks from videos using AI inpainting.

How it works

1. Detection: Temporal variance analysis identifies static overlay regions (watermarks). Falls back to YOLOv8n watermark detector if variance analysis fails.
2. Inpainting: LaMa (Large Mask Inpainting) fills in detected regions frame-by-frame.
3. Encoding: FFmpeg pipe-based I/O keeps memory usage constant regardless of video length.

Requirements

• Python 3.9+
• FFmpeg installed and on PATH

Install

pip install -r requirements.txt

Usage

Automatic detection (recommended)

python dewatermark.py -i input.mp4 -o output.mp4

Manual mask

python dewatermark.py -i input.mp4 -o output.mp4 -m mask.png

Manual region (X,Y,W,H)

python dewatermark.py -i input.mp4 -o output.mp4 --region 50,20,200,60

Options

Flag	Default	Description
-i	required	Input video path
-o	required	Output video path
-m	—	Mask image (white = watermark)
--region	—	Bounding box as X,Y,W,H
--model-dir	~/.cache/dewatermark	Model download directory
--feather	5	Gaussian blur on mask edges (px)
--crf	18	Output quality (lower = better)
--preset	medium	FFmpeg encoding preset
--sample-frames	30	Frames for variance analysis
--variance-threshold	auto	Variance sensitivity (0-1, default: Otsu)

Models

Models are downloaded automatically on first run:

• LaMa (Carve/LaMa-ONNX) — ~208MB, inpainting
• YOLOv8n (qfisch/yolov8n-watermark-detection) — ~12MB, watermark detection (fallback only)