This repository provides the official implementation for "The Stable Signature," a method for embedding robust watermarks within Latent Diffusion Models (LDMs). It enables users to embed and extract hidden messages in image generation pipelines, offering a way to trace or protect AI-generated content.

How It Works

The Stable Signature embeds watermarks by fine-tuning the decoder of an LDM. This approach modifies the decoder's weights to encode a binary message, which can then be extracted from generated images. The method prioritizes robustness against common image manipulations while maintaining high perceptual quality.

Quick Start & Requirements

• Installation: Clone the repository and install dependencies using conda and pip.
• Core Dependencies: Python 3.8, PyTorch 1.12.0, CUDA 11.3.
• Data: Requires a dataset (e.g., COCO) for fine-tuning the LDM decoder.
• Models: Downloadable pre-trained watermark extractor models (.pth and .torchscript.pt formats) are provided.
• Resources: Setup involves downloading models and potentially fine-tuning on a custom dataset.
• Links: Webpage, arXiv, Blog, Demo

Highlighted Details

• Fine-tunes LDM decoders for watermark embedding.
• Provides extractor models with varying robustness (e.g., against blur and rotations).
• Includes scripts for fine-tuning, generation with watermarked weights, and evaluation.
• Supports integration with Stability AI's Stable Diffusion codebase and the Diffusers library.

Maintenance & Community

• Developed by Meta AI (facebookresearch).
• The project is based on Stability AI's stablediffusion and PerceptualSimilarity repositories.

Licensing & Compatibility

• License: CC-BY-NC for the majority of the project. src/ldm and src/taming are MIT licensed.
• Compatibility: The CC-BY-NC license restricts commercial use.

Limitations & Caveats

The project's primary license (CC-BY-NC) prohibits commercial use. Integration with the Diffusers library is noted as "Still WIP" and may require future updates.