Introducing SynthID: A Tool for Watermarking and Identifying AI-Generated Images
AI-generated images created by ImagenAI are becoming more popular every day. But how can we better identify them, especially when they look so realistic?
Today, in partnership with Google Cloud, we’re launching a beta version of SynthID, a tool for watermarking and identifying AI-generated images. This tool embeds a digital watermark directly into the pixels of an image, making it undetectable to the human eye while still detectable for identification.
SynthID is being released to a limited number of Vertex AI customers using Imagen, our text-to-image model that creates photorealistic images based on input text. Generative AI technologies are rapidly advancing, making it difficult to distinguish between images created by an AI system and those that are not.
The ability to identify AI-generated content is critical in preventing the spread of misinformation. SynthID aims to empower users with knowledge of when they’re interacting with AI-generated media.
This new tool generates an imperceptible digital watermark for AI-generated images. Google Cloud is the first cloud provider to offer a tool for creating AI-generated images responsibly and confidently identifying them. Developed by Google DeepMind and refined in partnership with Google Research, SynthID aligns with our commitment to developing and deploying responsible AI.
Traditional watermarks are not effective for identifying AI-generated images, as they can easily be edited out. SynthID solves this problem by creating a watermark that remains detectable even after image manipulations such as cropping, resizing, and adding filters. This technology allows users to work with AI-generated content responsibly without compromising image quality.
The watermarking process uses two deep learning models that have been trained on a diverse set of images. These models work together to optimize the imperceptibility of the watermark while still accurately identifying watermarked content.
SynthID offers a robust and scalable approach to watermarking and identifying AI-generated images. It provides three confidence levels for interpreting the results of watermark identification, helping users assess the likelihood of an image being created by Imagen.
In addition to watermarking, metadata is another widely used method for identifying digital content. However, metadata can be manually removed or lost during file editing. SynthID’s watermark is embedded in the pixels of an image, making it compatible with other identification approaches that rely on metadata, even if the metadata is lost.
As generative models advance and expand to other mediums, it’s important to develop adaptable and robust solutions. SynthID is just one piece of the puzzle in building AI-generated content responsibly. We look forward to gathering user feedback, enhancing capabilities, and exploring new features to ensure that SynthID evolves with the needs of creators and users.
In the future, SynthID could be integrated into more Google products and made available to third parties, further empowering individuals and organizations to work with AI-generated content responsibly.
Please note that the model used in this blog may differ from the model used on Imagen and Vertex AI.
Acknowledgements:
This project was led by Sven Gowal and Pushmeet Kohli, with contributions from Rudy Bunel, Jamie Hayes, Sylvestre-Alvise Rebuffi, Florian Stimberg, David Stutz, Meghana Thotakuri, and many others at Google DeepMind and Google. We would like to thank Nidhi Vyas and Zahra Ahmed for product delivery, Chris Gamble for project initiation, and Ian Goodfellow, Chris Bregler, and Oriol Vinyals for their advice.