Google DeepMind has introduced SynthID, a system that adds a digital watermark directly into the pixels of AI-generated images, making it invisible to the human eye but detectable by the system.
The technology behind SynthID relies on two deep learning models – one for watermarking and another for identification – that were trained together using a diverse set of images. Currently, SynthID is only available to a limited number of Google Vertex AI platform customers who use the Imagen AI image generator. However, the DeepMind team plans to expand its integration into more products, explore additional features and enhance its capabilities based on user feedback.
Unlike traditional watermarks that can be easily edited out, SynthID’s AI watermark doesn’t compromise image quality and remains detectable even after modifications like applying filters, changing colours or saving with various compression schemes. The tool offers three confidence levels for interpreting watermark identification results, helping users determine if a portion of the image was generated by Imagen.
While SynthID is an important step in watermarking AI-generated content, the DeepMind team acknowledges that it may not be foolproof against extreme image manipulations. They also anticipate that the technology could evolve to cover various AI uses beyond imagery, including audio, video and text.
ADD WORKGROUP