Content Seal is a comprehensive framework for invisible, robust watermarking across all modalities — audio, image, video, and text. It spans the entire generative AI lifecycle, from training data and inference to generated media, providing state-of-the-art tools for content provenance and authentication.
Watermarks applied after content generation by any model or system — model-agnostic and universal across all content types.
Content Seal Image is deployed at scale for Muse Image with a custom proprietary implementation. We also provide open-source versions of our research models for images and video, readily available to download.
| Research | Links |
|---|---|
| Pixel Seal: Adversarial-Only Training for Invisible Image and Video Watermarking Flagship image & video watermarking model, SOTA in robustness and imperceptibility, built with a more stable adversarial-only training paradigm. |
  |
| We Can Hide More Bits: The Unused Watermarking Capacity in Theory and in Practice Bigger model with 4× capacity boost to 1024 bits while preserving quality and robustness. |
 |
| Video Seal: Open and Efficient Video Watermarking Extension of image watermarking models to video, resilient to editing and video codecs. |
 |
| Watermark Anything with Localized Messages Embed (possibly multiple) localized watermarks into images; survives inpainting and splicing attacks. |
|
| Geometric Image Synchronization with Deep Watermarking Robust image synchronization, enabling reversal of geometric transformations applied to an image. |
|
| Research | Links |
|---|---|
| Proactive Detection of Voice Cloning with Localized Watermarking Localized audio watermarking with sample-level detection and streaming support for real-time applications. |
|
| Research | Links |
|---|---|
| How Good is Post-Hoc Watermarking With Language Model Rephrasing? Comprehensive evaluation framework for post-hoc text watermarking with LLM rephrasing. |
|
Watermarks embedded during content generation by modifying model behavior or latent representations.
| Research | Links |
|---|---|
| TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection SOTA LLM watermark with dual-key Gumbel-max sampling, entropy-weighted scoring, and multi-region localization. Distortion-free, preserves reasoning and benchmark performance, detectable when diluted in human text, radioactive through distillation. |
|
| Research | Links |
|---|---|
| Learning to Watermark in the Latent Space of Generative Models Unified latent-space watermarking that enables 20× speedup over pixel methods and secures open-source models via in-model distillation. |
|
| The Stable Signature: Rooting Watermarks in Latent Diffusion Models Roots the watermark in the model's latent decoder for tracing the outputs of latent generative models. |
|
| Watermarking Autoregressive Image Generation Watermarking for autoregressive image generation models. |
|
| Research | Links |
|---|---|
| Watermarking Makes Language Models Radioactive Detects if a language model was trained on synthetic text by finding weak residuals of watermark signals in fine-tuned LLMs — high-confidence even when as little as 5% of training text is watermarked. |
|
| Detecting Benchmark Contamination Through Watermarking Watermarks benchmarks before release to detect if models were trained on test sets, using theoretically grounded statistical tests while preserving benchmark utility. |
|
Research on adversarial attacks and defenses for watermarking systems through red teaming.
| Research | Links |
|---|---|
| Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models Black-box watermark forging using image preference models for red-teaming watermarking systems. |
|
The code is licensed under an MIT license.