Svetozar1993/videoseal-1.0

VideoSeal 1.0

Neural network-based invisible video watermarking model from Meta's VideoSeal.

Model Details

• Architecture: UNet embedder (Y-channel) + ConvNeXt Tiny detector
• Message length: 256 bits
• Processing resolution: 256x256
• Blending: Additive (default strength: 0.2)
• Attenuation: JND (Just Noticeable Difference)

Usage

pip install videoseal torch torchvision ffmpeg-python tqdm

# Embed watermark (auto-generates unique 256-bit message)
python video_watermark.py embed -i input.mp4 -o watermarked.mp4

# Extract watermark
python video_watermark.py extract -i watermarked.mp4

# Verify against registry
python video_watermark.py verify -i suspect.mp4

Robustness

Scenario	Bit Accuracy
Direct extraction	~100%
Re-encoded at CRF 30	~98%
Resolution scaling	Strong
Multiple re-encodes	Strong

Files

• y_256b_img.pth — PyTorch checkpoint (~228MB)

Source

• VideoSeal GitHub
• Our CLI tool