Meta has announced Meta Video Seal, a new open-source tool designed to watermark AI-generated videos. Video Seal embeds invisible yet robust watermarks that persist through edits, compression, and sharing, making it possible to trace and authenticate content.
The release aims to address growing concerns over the misuse of generative AI, including the rise of deepfakes and manipulated content.
The development of Video Seal comes at a time when the proliferation of generative AI has introduced new challenges for content moderation. Research found a tenfold (10x) increase in the number of deepfakes detected globally across all industries from 2022 to 2023, contributing to 7% of global fraud cases, ranging from identity theft to elaborate social engineering.
With the release of OpenAI’s Sora video generator this week, such watermarking solutions will have to play an increasing role to help humans distinguish AI generated videos from real footage.
The Problem Video Seal Solves
Deepfake technology, powered by generative AI models like DALL·E and Stable Diffusion, has enabled the creation of hyper-realistic but deceptive content.
While these tools offer creative possibilities, they also pose risks, such as the spread of misinformation and fraudulent activities. Video watermarking offers a solution by embedding identifiers within videos to verify their authenticity.
Traditional video watermarking techniques often fall short of addressing modern challenges. Video transformations such as cropping, compression, and re-encoding can distort or remove watermarks, rendering them ineffective. Meta’s Video Seal introduces new methods to ensure resilience against such distortions.
How Video Seal Works
Meta Video Seal employs innovative neural watermarking techniques that make it both efficient and robust. One of its defining features is temporal watermark propagation, a process that embeds a watermark in key video frames and propagates it to neighboring frames.
This approach reduces computational load while ensuring the invisible watermark remains intact even after the video undergoes common transformations.
The tool also leverages differentiable augmentations, transformations applied during the training phase that simulate real-world video distortions, to simulate real-world distortions such as compression and cropping. This ensures the watermark’s resilience under challenging conditions.
Meta’s research team explains, “Video Seal eliminates the need to watermark every frame in a video by leveraging temporal watermark propagation, enabling fast inference times while maintaining robustness to video compression and geometric distortions.”
Another critical aspect is the multistage training process. Video Seal begins with image pre-training, transitioning to video-specific training with fine-tuning stages. This hybrid approach combines the efficiency of image-based models with the adaptability needed for video content.
Video Seal has been tested against industry benchmarks like MBRS, TrustMark and WAM and demonstrated superior performance under transformations such as H.264 compression and geometric distortions. This makes it one of the most robust watermarking frameworks currently available.
To encourage further collaboration, Meta has launched the Meta Omni Seal Bench, a public leaderboard for comparing watermarking techniques. The company also plans to host a workshop at the International Conference on Learning Representations (ICLR) in 2025 to foster dialogue within the academic and industrial communities.
“We are publicly releasing the Video Seal model under a permissive license, along with a research paper, training code, and inference code.” Meta states.
Other AI Watermarking Solutions
Meta is not alone in its efforts to address the challenges posed by generative AI. Competitors like DeepMind and Microsoft have also introduced watermarking tools, highlighting an industry-wide recognition of the issue.
Google DeepMind’s SynthID, for instance, uses similar neural watermarking techniques for AI-generated images, videos, and audio. However, Video Seal’s open-source nature sets it apart, allowing researchers and developers to freely access its code and contribute to its improvement.
Meta Video Seal also focuses exclusively on video content, embedding imperceptible watermarks into individual frames and offering optional hidden messages for tracing video origins.
In contrast, Google’s SynthID applies a broader strategy, embedding watermarks into text, images, audio, and video. Both technologies demonstrate resilience against common edits like cropping, compression, and filtering.
Applications and Challenges
While Video Seal is primarily aimed at combating misinformation, its applications extend far beyond. Industries like media and entertainment could use the technology to prevent piracy and verify content authenticity. For example, Hollywood studios could embed watermarks into movie files to track unauthorized copies.
However, there are limitations. Heavily compressed videos or those subjected to extensive edits may degrade the watermark’s signal. The research team acknowledges, that although Video Seal demonstrates high robustness against common transformations, including compression and geometric edits, its watermarking may degrade or become irretrievable under extreme conditions like heavy compression or substantial modifications.
Another challenge lies in adoption. Although Meta has made Video Seal open source, widespread industry adoption depends on compatibility with existing workflows and a willingness to move away from proprietary systems.
The introduction of Video Seal reflects Meta’s stated commitment to responsible AI development. Earlier this year, the company reported that less than 1% of election-related misinformation detected on its platforms was AI-generated, underscoring its proactive approach to content moderation. This aligns with Meta’s broader strategy to mitigate AI misuse while fostering innovation.
By releasing Video Seal as an open-source tool, Meta has taken an important step toward addressing the challenges posed by generative AI. Its ability to embed imperceptible yet resilient watermarks into videos not only combats misuse but also strengthens trust in digital content. As industries grapple with the implications of AI, tools like Video Seal could become essential for maintaining transparency and authenticity.
