Google is launching the second version of Imagen, a powerful AI model that can generate and modify images based on a text prompt. The new model, Imagen 2, is now available to Google Cloud customers using Vertex AI who have been approved for access.
However, Google has not disclosed the specific data used to train the new model. Additionally, there is no option for creators who may have unintentionally contributed to the dataset to opt out or seek compensation.
Imagen 2, which was quietly previewed at Google’s I/O conference in May, was developed using technology from Google DeepMind, the company’s leading AI lab. According to Google, Imagen 2 has significantly improved image quality compared to the first-generation Imagen. It also introduces new capabilities such as rendering text and logos.
During a press briefing, Google Cloud CEO Thomas Kurian mentioned that Imagen 2 allows users to create images with a text overlay, which can be useful for advertising purposes.
The addition of text and logo generation brings Imagen 2 in line with other prominent image-generating models, such as OpenAI’s DALL-E 3 and Amazon’s Titan Image Generator. However, Imagen 2 can differentiate itself in two key aspects. It can render text in multiple languages, including Chinese, Hindi, Japanese, Korean, Portuguese, English, and Spanish, with more languages expected to be supported in 2024. Moreover, it can overlay logos onto existing images.
Vishy Tirumalasetty, head of generative media products at Google, explained in a blog post that Imagen 2 can generate emblems, lettermarks, abstract logos, and overlay these logos onto products, clothing, business cards, and other surfaces.
Google has utilized advanced training and modeling techniques in Imagen 2, enabling it to process more descriptive prompts and provide detailed responses to questions about elements in an image. The model’s multilingual understanding allows it to translate a prompt in one language to an output in another language.
Moreover, Imagen 2 leverages SynthID, a technique developed by DeepMind, to apply invisible watermarks to the images it creates. These watermarks are designed to be resilient to image edits including compression, filters, and color adjustments. Google has also introduced a tool for detecting these watermarks, addressing concerns over the growing volume of AI-generated disinformation.
However, Google has not revealed the specific data used to train Imagen 2, which has raised legal questions regarding the use of publicly available and copyrighted data for commercial purposes. The company has taken a cautious approach in response, as ongoing lawsuits and legal debates continue.
While some companies offer creators the option to opt out of training datasets or provide compensation, Google and several other major players in the industry have not implemented such mechanisms. Instead, Google offers an indemnification policy to protect eligible Vertex AI customers from copyright claims related to the use of training data and Imagen 2 outputs.
Addressing concerns around regurgitation, where generative models produce copies of training examples, Google hopes that its expanded policy will assuage worries about intellectual property. Nevertheless, concerns over compensation for creators remain unaddressed at this time.
