Google is enhancing its image-generation technology to stay competitive with industry peers.
During the I/O developer conference held in Mountain View on Tuesday, Google introduced Imagen 3, the latest iteration in its Imagen generative AI series.
Demis Hassabis, CEO of DeepMind—Google’s AI research arm—explained that Imagen 3 has a more precise understanding of text prompts compared to its predecessor, Imagen 2, and delivers more “creative and detailed” image generation. Additionally, the model has been refined to produce fewer “distracting artifacts” and errors.
“This is also our most capable model to date for rendering text, which has been a difficult task for image-generation models,” said Hassabis.
To address concerns about the potential misuse for creating deepfakes, Google has incorporated SynthID, a technique developed by DeepMind that embeds invisible, cryptographic watermarks into the media.
Private preview sign-ups for Imagen 3 are currently available through Google’s ImageFX tool, and Google plans to soon make the model accessible to developers and enterprise clients via Vertex AI, Google’s generative AI development platform.
Google generally keeps the details of its data sources for AI model training under wraps, and this instance is no different. The underlying reason is that much of the training data is sourced from publicly accessible sites, repositories, and datasets across the internet. A portion of this data, namely copyrighted material scraped without the explicit consent of content creators, has led to a number of IP-related legal challenges.
Google’s web publisher tools enable site owners to block the company from scraping data, including images and videos, from their sites. However, Google does not provide a comprehensive “opt-out” mechanism and, unlike several of its competitors, has not pledged to remunerate rights holders for their inadvertent contributions to its training datasets.
The lack of transparency is not unexpected but, given Google’s extensive resources, it remains a disappointment.
