**New Tool Helps Watermark and Identify Synthetic Images Created by Imagen**
AI-generated images are gaining popularity, but it can be challenging to identify them, especially since they look so realistic. That’s why we’re introducing SynthID, a new tool developed in partnership with Google Cloud. SynthID allows users to watermark and identify AI-generated images by embedding a digital watermark directly into the pixels of the image, making it undetectable to the human eye but detectable for identification.
SynthID is currently available in beta version and is being offered to a limited number of Vertex AI customers who use Imagen, our text-to-image model. With the growing advancements in generative AI technologies, it’s becoming increasingly difficult to distinguish computer-generated images, also known as “synthetic imagery,” from those created by humans. This poses risks as it enables creators to spread false information, intentionally or unintentionally. Identifying AI-generated content is crucial for empowering people with knowledge and preventing the spread of misinformation.
Google Cloud is the first cloud provider to offer a tool for creating AI-generated images responsibly and identifying them with confidence. SynthID was developed by Google DeepMind and refined in collaboration with Google Research. While not foolproof against extreme image manipulations, SynthID offers a promising technical approach to enable responsible use of AI-generated content. The tool may also evolve to encompass other AI models and modalities like audio, video, and text.
**A New Type of Watermark for AI Images**
Traditional watermarks are insufficient for identifying AI-generated images as they can easily be edited out. SynthID addresses this challenge by striking a balance between imperceptibility and robustness. It ensures that the watermark remains detectable even after various modifications and commonly used lossy compression schemes for images like JPEGs.
SynthID employs two deep learning models for watermarking and identification, trained together on a diverse set of images. The models are optimized to correctly identify watermarked content and enhance imperceptibility by visually aligning the watermark with the original content.
**Robust and Scalable Approach**
SynthID empowers Vertex AI customers to create AI-generated images responsibly while confidently identifying them. Although the technology isn’t flawless, internal testing shows its accuracy against many common image manipulations.
SynthID offers a combined approach:
1. Watermarking: SynthID adds an imperceptible watermark to synthetic images produced by Imagen.
2. Identification: By scanning an image for its digital watermark, SynthID can determine the likelihood of it being created by Imagen.
The tool provides three confidence levels for interpreting watermark identification results, allowing users to assess the origin of the image.
Metadata is a widely used method for identifying digital content, but it can be easily removed or lost during editing. SynthID’s embedded watermark remains detectable even when metadata is lost, making it compatible with other image identification approaches based on metadata.
**What’s Next?**
We are committed to developing safe, secure, and trustworthy approaches for building AI-generated content. As generative models advance and expand to other mediums, robust and adaptable solutions are needed. We believe that SynthID can work in conjunction with a range of solutions for creators and users across society. We will continue to gather feedback, enhance SynthID’s capabilities, and explore new features. Furthermore, we aim to integrate SynthID into more Google products and make it available to third parties, enabling responsible use of AI-generated content.
(*Note: The synthetic image model mentioned in this blog may differ from the one used on Imagen and Vertex AI.*)