SynthID, is a watermarking and identification tool for generative art. The company says the technology embeds a digital watermark, invisible to the human eye, directly onto an image’s pixels.
In a bold move towards bolstering trust in the ever-evolving landscape of AI-generated content, Google has announce the beta launch of SynthID—a groundbreaking tool developed by Google DeepMind, refined in collaboration with Google Research. SynthID serves as a cutting-edge solution to the pervasive challenge of misinformation by allowing users to embed imperceptible digital watermarks directly into AI-generated images and audio.
The imperative need to identify AI-generated content underscores the significance of SynthID in the realm of enterprise solutions. While not a solution for misinformation, this tool represents a crucial step forward in addressing AI safety concerns and promoting responsible engagement with synthesized content.
How SynthID Works
SynthID employs two deep learning models—one for watermarking and another for identification:
Watermarking: The embedded watermarking technology seamlessly integrates a digital watermark into AI-generated content, optimizing imperceptibility by precisely aligning the watermark with the original material.
Identification: SynthID scans AI-generated images or audio for its digital watermark, enabling users to ascertain whether the content, or a portion of it, was produced using their AI models.
SynthID in Action: AI-Generated Music
In a significant expansion in November 2023, SynthID now encompasses the watermarking and identification of AI-generated music and audio. This functionality is initially deployed through Lyria, Google’s state-of-the-art AI music generation model. SynthID discreetly embeds digital watermarks directly into the audio waveform of AI-generated content, ensuring imperceptibility to the human ear.
During the watermarking process, SynthID converts the audio wave into a spectrogram—a two-dimensional visualization of the sound spectrum over time. The digital watermark is seamlessly integrated into the spectrogram, and the process is reversed to maintain inaudibility while preserving the listening experience. SynthID’s robust watermark remains detectable even after common modifications like noise additions, MP3 compression, or alterations in speed.
SynthID for AI-Generated Images: An Enterprise Perspective
In catering to enterprise needs, SynthID is made available to a select group of Vertex AI customers utilizing the Imagen suite—a cutting-edge text-to-image model. This suite enables the creation of photorealistic images from input text.
For AI-generated images, SynthID embeds digital watermarks directly into the pixels, ensuring that is unnoticable to the human eye. Designed to preserve image quality, the watermark remains detectable even after common modifications such as adding filters, changing colors, or utilizing lossy compression schemes.
SynthID provides users with three confidence levels for interpreting results, offering a robust solution for identifying AI-generated content. Whether applied to music or images, SynthID plays a pivotal role in fortifying enterprise solutions and empowering organizations to responsibly navigate the realm of AI-generated content.
If you have more questions about how to integrate SynthID into your products, reach out to us and let us know how we an help you.