Video Seal introduces a novel open-source framework for neural video watermarking, combining a jointly trained embedder and extractor to ensure robustness and efficiency. The method leverages multistage training—image pre-training, hybrid post-training, and extractor fine-tuning—alongside temporal watermark propagation to avoid per-frame embedding. Differentiable augmentations, including common video codecs, are applied during training to simulate real-world transformations and improve resilience.
The accompanying codebase provides end-to-end pipelines for training and inference. Key assets include TorchScript models, training scripts (train.py), inference tools (inference_av.py, inference_streaming.py), evaluation modules (videoseal/evals/full.py), and Jupyter notebooks for interactive exploration. Pre-trained checkpoints and a live demo endpoint are also available to facilitate rapid experimentation.