THU-BPM/MarkDiffusion

What it solves

MarkDiffusion addresses the need for authenticity and origin verification in media generated by latent diffusion models (LDMs). It provides a standardized way to embed, detect, and evaluate watermarks in AI-generated images and videos to prevent misuse and track provenance.

How it works

The toolkit provides a unified Python framework that integrates multiple watermarking algorithms into a single interface. It supports two main categories of methods: pattern-based (which embed specific fingerprints) and key-based (which use cryptographic keys). The system includes a visualization suite to show how watermarks are embedded and extracted, and an evaluation module that tests the watermarks against various "attacks" (like cropping, compression, or noise) to measure their robustness and impact on the final media quality.

Who it’s for

This toolkit is designed for AI researchers and developers working with generative models who need to implement watermarking for security, copyright protection, or authenticity verification.

Highlights

• Broad Algorithm Support: Implements 11 state-of-the-art watermarking algorithms for both images and videos.
• Comprehensive Evaluation: Includes 33 tools to measure detectability, robustness, and output quality.
• Robustness Testing: Features specialized attack tools including JPEG compression, Gaussian blur, and advanced regeneration attacks like Diffusion Purification.
• Automated Pipelines: Provides 8 automated evaluation pipelines for streamlined quality and detection analysis.