How It Works
VideoSeal is a neural watermarking framework developed by Meta AI Research that embeds and extracts imperceptible marks in images and videos. The system relies on deep learning models trained to balance robustness against common transformations while maintaining visual imperceptibility. Three core models—VideoSeal, PixelSeal, and ChunkySeal—address different trade-offs between watermark capacity, robustness, and imperceptibility.
The embedding process modifies pixel values in a way that is statistically detectable but visually undetectable. ChunkySeal increases watermark capacity to 1024 bits while preserving image quality, demonstrating that current watermarking limits are not yet exhausted.
Just Noticeable Difference (JND) masking is a key technique used in VideoSeal and PixelSeal. This method generates a heatmap where watermark strength is attenuated in smooth regions and amplified in textured areas, reducing visibility without compromising robustness. The framework supports adjustable watermark strength via the scalingw parameter, allowing users to trade off between robustness and imperceptibility based on specific use cases.
Models are implemented in PyTorch and optimized for inference efficiency. VideoSeal v1.0, released in March 2025, supports 256-bit watermarks with enhanced robustness, while earlier versions like VideoSeal v0.0 provided a 96-bit baseline. All models are released under the MIT license, ensuring permissive use in research and commercial applications.
Use Cases
VideoSeal is designed for provenance verification and content authentication, where the ability to embed and detect watermarks without altering visual quality is critical.
PixelSeal’s state-of-the-art imperceptibility and robustness trade-off positions it as a strong candidate for high-security applications, such as secure document distribution or digital asset tracking.
The open-source nature of VideoSeal facilitates integration into existing pipelines. The inclusion of evaluation tools, such as VMAF (Video Multi-Method Assessment Fusion), allows users to quantify imperceptibility and robustness against standardized benchmarks.
Model Variants and Trade-offs
- VideoSeal v1.0: A stable 256-bit model designed for video watermarking with temporal consistency. It balances robustness and imperceptibility for general-purpose use, including copyright protection and content tracking.
- PixelSeal: Achieves state-of-the-art performance in robustness and imperceptibility through adversarial-only training and JND-based attenuation.
Each model is accompanied by a model card and pre-trained checkpoints, enabling immediate deployment without extensive training. The framework also supports baseline comparisons against third-party models, though these must be manually downloaded.
Implementation and Performance
VideoSeal is implemented in Python using PyTorch, with dependencies including TorchVision, TorchAudio, and CUDA 12.1 for accelerated inference. The library supports both image and video watermarking, with specialized scripts for streaming embedding and extraction to handle long videos efficiently.
- Just Noticeable Difference (JND) Masking: This technique dynamically adjusts watermark strength based on local image texture, reducing visibility in smooth regions while preserving robustness in textured areas.
- Adjustable Watermark Strength: The scalingw parameter allows users to modify global watermark strength, trading off between robustness and imperceptibility. For example, increasing scalingw enhances resistance to attacks but may make the watermark more perceptible.
Evaluation and Benchmarking
VideoSeal includes comprehensive evaluation tools to assess imperceptibility and robustness. The full.py script in the videoseal/evals/ directory generates metrics such as VMAF scores and robustness against augmentations like compression, cropping, or noise addition. These metrics are saved to a CSV file for further analysis.
For video evaluation, the framework recommends using the SA-V dataset, a high-resolution benchmark for video watermarking. Pre-computed metrics for VideoSeal v1.0 on this dataset are available as a reference, ensuring reproducibility and comparability with other methods.
Training and Customization
The repository provides training code and documentation to reproduce or extend the models. Users can train custom watermarking models by following the instructions in docs/training.md, which covers data preparation, training commands, and checkpoint management. Training is recommended to start with image models before extending to video.
- Three-Stage Training Schedule: Addresses resolution scaling challenges by progressively increasing input size during training.
- JND-Based Attenuation: Integrated into the training pipeline to optimize watermark visibility across different image regions.
Dependencies and Compatibility
VideoSeal requires Python 3.10 and specific versions of PyTorch (2.4.0), TorchVision (0.19.0), and TorchAudio (2.4.0). CUDA 12.1 is recommended for GPU acceleration. Additional dependencies like decord are optional but recommended for training, as they improve video processing efficiency.
Audio-Visual Watermarking
VideoSeal extends its capabilities to audio-visual watermarking through the inferenceav.py script. This tool embeds and detects watermarks in both audio and video streams from a single file. However, it loads the entire video into memory, making it unsuitable for very long or high-resolution videos. For such cases, the streaming-based approach is preferred.
Open-Source and Community Support
The project is maintained by researchers from Meta AI, including Pierre Fernandez, Hady Elsahar, Tomas Soucek, Sylvestre Rebuffi, and Alex Mourachko. Contributions are welcome, with guidelines provided in the repository’s CONTRIBUTING.md file.
Related projects include AudioSeal for audio watermarking and Watermark-Anything, a broader framework for multi-modal watermarking applications.