SynthID, a tool for watermarking and identifying AI-generated photos, has now released its beta edition in collaboration with Google Cloud. This technology embeds a pixel-level digital watermark onto an image for verification, yet is invisible to the naked eye.

Imagen, one of the most recent text-to-image models, is being used to deliver SynthID to a select group of Vertex AI customers. Imagen takes in text and outputs a photorealistic image.

Generative AI technologies are changing quickly, and it is getting harder to tell the difference between computer-generated images, also called "synthetic imagery," and images not made by an AI system. Although generative AI can open up a lot of creative potential, it also comes with new risks, like letting people spread fake information on purpose or by accident. People need to recognize AI-generated material to know when they're interacting with it and to help stop the spread of false information.

Google Cloud

Google Cloud is the first to offer a solution for safely making and identifying AI-generated photographs. This technology, developed by Google DeepMind and refined in collaboration with Google Research, is based on the approach to building and implementing responsible AI.

SynthID is not impervious to extreme image manipulation, but it provides a viable technical method for empowering people and organizations to operate ethically with AI-generated material. This tool may evolve with AI models and modalities other than pictures, such as voice, video, and text.

Watermarks

Watermarks are designs that can be overlaid on photographs to help them be identified. They've changed throughout history, from physical imprints on paper to the transparent language and symbols in digital pictures today. Traditional watermarks are insufficient for distinguishing AI-generated images because they are frequently applied as a stamp on an image and can be readily removed. For example, discrete watermarks in an image's corner can be edited out using basic editing techniques.

It is challenging to distinguish between imperceptibility and resistance against image alteration. Obvious watermarks, such as a name or brand on top of an image, cause aesthetic concerns for artistic or economic purposes. Similarly, some previously generated invisible watermarks can be removed using basic editing procedures such as resizing.

SynthID

The researchers created SynthID to preserve image quality while allowing the watermark to be detected after alterations such as adding filters, changing colours, and storing with various lossy compression algorithms — most typically used for JPEGs. 

SynthID employs two deep learning models, one for watermarking and the other for identifying, that have been trained on a broad set of photos. The combined model is tuned to achieve various goals, including correctly recognizing watermarked information and enhancing imperceptibility by aesthetically aligning the watermark with the original content.

Conclusion

SynthID technology may collaborate with various solutions for producers and users throughout society. 

Furthermore, SynthID has the potential to be developed for use across various AI models, and the researchers are thrilled about the prospect of integrating it into more Google products and making it available to third parties soon – empowering people and organizations to operate responsibly with AI-generated content.