Reproduce “Video Seal: Open and Efficient Video Watermarking” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Pierre Fernandez, Hady Elsahar, I. Zeki Yalniz, Alexandre Mourachko
Meta FAIR researchers developed Video Seal, an open and efficient neural video watermarking framework capable of imperceptibly embedding and extracting messages in high-resolution, long videos. This framework demonstrates competitive robustness against various video-specific distortions, including compression and geometric transformations, while maintaining high visual quality and computational efficiency.
Key insights
- Hybrid training, specifically image pre-training followed by a mix of images and videos with differentiable compression augmentations, is crucial for achieving high robustness in video watermarking.
- Separately fine-tuning the extractor with a frozen embedder significantly improves robustness against various attacks without degrading the watermark's imperceptibility.
- Temporal watermark propagation effectively balances embedding speed with robustness, demonstrating that it is not necessary to fully embed a watermark on every single frame of a video.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.