State-of-the-Art, Open Source Invisible Watermarking
Introducing Meta Seal
State-of-the-Art, Open Source Invisible Watermarking
A state-of-the-art, open-source framework for invisible, robust watermarking across all modalities audio, image, video, and text. This suite spans the entire generative lifecycle, from training data and inference to generated media.
Read the Research Papers Get the Models
Research Overview
Watermarking across all modalities
Meta Seal offers state-of-the-art, open-source tools for embedding robust and imperceptible watermarks across images, video, audio, text, and generative AI models. As AI-generated content becomes increasingly sophisticated, provenance tracking and authentication mechanisms are more critical than ever.
Post-hoc WM Generative WM Dataset WM Security
Post-hoc watermarking
Watermarks applied after content generation by any model or system. This model-agnostic approach works universally across images, video, audio, and text—no matter how the content was created. Designed for protecting existing content and third-party generated media.
##### Pixel Seal
Flagship image & video watermarking model SOTA in terms of robustness and imperceptibility, which is built with a better and more stable adversarial-only training paradigm.
##### Chunky Seal
Bigger model with 4× capacity boost to 1024 bits while preserving quality and robustness.
##### Video Seal
Extension of image watermarking models to video, resilient to editing and video codecs.
##### WAM (Watermark Anything Model)
Embed (possibly multiple) localized watermarks into images, survives inpainting and splicing attacks.
##### Sync Seal
Watermarking models for robust image synchronization, enabling to revert geometric transformations applied to image.
Show 2 more papers ↓
##### Audio Seal & Audio Seal Streaming
Localized audio watermarking with sample-level detection and streaming support for real-time applications.
##### Text Seal
Comprehensive evaluation framework for post-hoc text watermarking with LLM rephrasing.
In-model & generation-time watermarking
Watermarks embedded during the generation process by modifying the model's latent representations or embedding the watermark directly in the model weights. This native approach enables seamless integration with diffusion models, LLMs, and other generative systems for real-time watermarking as content is created.
Unified latent space watermarking that enables 20× speedup over pixel methods and secures open-source models via in-model distillation.
Roots the watermark in the model's latent decoder for tracing the outputs of latent generative models.
Watermarking for autoregressive image generation models.
Dataset watermarking
Watermarks embedded into training datasets to track data provenance and usage. This technique enables dataset creators to verify if their data was used to train specific models, providing accountability and attribution in the AI training pipeline.
Designed to detect if a language model was trained on synthetic text by detecting weak residuals of watermark signals in fine-tuned LLMs, with high confidence detection even when as little as 5% of training text is watermarked.
Watermarks benchmarks before release to detect if models were trained on test sets, using theoretically grounded statistical tests to identify contamination while preserving benchmark utility.
Watermark security
Research on adversarial attacks and defenses for watermarking systems. Red teaming efforts explore vulnerabilities like watermark removal, forgery, and spoofing attacks—while developing robust countermeasures to help watermarks remain secure and reliable.
Black-box watermark forging using image preference models for red-teaming watermarking systems.
Try it yourself
Get the Models
Access our official research implementations. Open weights, MIT Licensed, and designed for immediate scientific exploration.
Video Seal Audio Seal Text Seal
←
Watermarking at Scale
Learn how Meta scaled invisible watermarking from research to production, addressing real-world challenges including originality protection on Instagram and detecting AI-generated videos. Read the engineering blog.
Seamless Interaction Dataset
Audio Seal and Video Seal watermark 4,000+ hours of dyadic conversations with streaming capability and robust detection for AI-generated content. Explore dataset.
Llama AI Defenders
Audio watermarking and detection technology that provides industry-leading detection performance on accuracy, imperceptibility, and speed. Learn more.
GenAI Transparency
Meta uses invisible watermarking to enable transparency labels on AI-generated content across Facebook, Instagram, and Threads. Learn about AI-generated image labeling and transparency in ads.
→