THU-BPM/MarkDiffusion

[JMLR] MarkDiffusion: An Open-Source Toolkit for Generative Watermarking of Latent Diffusion Models

What it solves

MarkDiffusion addresses the challenge of verifying the authenticity and origin of media created by latent diffusion models. As generative AI becomes more prevalent, this toolkit provides a standardized way to embed and detect watermarks, helping researchers and developers ensure that generated images and videos can be reliably identified.

How it works

The toolkit offers a unified framework that integrates 11 state-of-the-art watermarking algorithms, categorized into pattern-based and key-based methods. Users can generate watermarked media directly through the library, visualize how these patterns are embedded, and use a comprehensive evaluation module to test the effectiveness of these watermarks against various "attacks" (such as cropping, noise, or compression) and assess their impact on output quality.

Who it’s for

It is designed for researchers, developers, and practitioners working with generative AI who need to implement, test, or compare watermarking techniques for diffusion models. It provides both a simple interface for quick integration and an extensive suite of tools for rigorous academic or technical evaluation.

Highlights

  • Supports 11 state-of-the-art watermarking algorithms for both images and videos.
  • Includes a robust evaluation module with 33 tools covering detectability, robustness, and image/video quality.
  • Features 8 automated evaluation pipelines for streamlined testing and benchmarking.
  • Provides visualization tools to help users understand how watermark patterns are applied and extracted.
  • Includes a comprehensive test suite with 672 unit tests and high code coverage to ensure reliability.

Related

  • Project
    facebookresearch/stable_signature

    A system for embedding invisible watermarks directly into the LDM decoder of generative AI models to automatically identify AI-generated images.

  • Dispatch
    AI Watermarking 101: Tools and Techniques

    Hugging Face provides a comprehensive overview of AI watermarking techniques across images, text, and audio to combat deepfakes and ensure content provenance.

  • Dispatch
    Visible Watermarking with Gradio

    Hugging Face has introduced a simple way to add visible watermarks to AI-generated images, videos, and text using the Gradio library to improve synthetic content transparency.

  • Project
    THU-BPM/MarkLLM

    MarkLLM is a Python toolkit that implements dozens of text‑watermarking methods for large language models, offers visualisation of how the watermarks work, and provides a full evaluation suite (detection, robustness, text‑quality metrics). It can be installed via pip, loads pretrained models from Hugging Face, and lets users generate and detect watermarks through a unified `AutoWatermark` API.

  • Project
    wiltodelta/remove-ai-watermarks

    A Python library/CLI for detecting and removing AI‑generated watermarks (visible logos, hidden pixel marks, and metadata) from images and videos you own, with optional GPU‑based diffusion for invisible marks.