Imagen 2 is the latest version of Google’s AI model that can convert text into images. Image quality and answers are improving, but details about how to train the model remain unknown.
Google unveiled its most powerful AI models to date, Gemini and Gemini Pro, a few days ago in response to OpenAI’s GPT-4. The next launch in line is Imagen 2. It is an AI model that can generate images based on text prompts. This updated version differs from the previous version of Imagen, provides better image quality and can provide more detailed answers. However, it is not known what data Google used to train this model.
Improve image quality and smarter
Imagen 2 was developed using technology from Google DeepMind: Google’s leading artificial intelligence lab. Imagen 2 is able to create and edit an image based on text prompts.
In addition to improving the quality of images, the AI model also introduces new options for displaying text and logos, such as image with text overlap. In addition, Imagen 2 is also capable of creating logos, characters, and abstract slogans. These logos can then be pasted onto products, clothing or business cards.
Finally, Imagen 2 is also smarter than its predecessor. This model is trained to better understand more descriptive and longitudinal evidence and provide detailed answers to questions about specific elements in the image. Here, Google emphasizes the multilingual understanding of Imagen 2, so it can output a claim in one language to another (eg a logo).
Training techniques for the model are unknown
Despite the improvements to the AI model, Google has not disclosed the data it used to train Imagen 2. As a result, creators who inadvertently contributed to the data set cannot opt out or seek compensation.
Intellectual property is a major concern when using generative AI. Some companies, like Stability AI and OpenAI, offer creators the option to opt out of training datasets. On the other hand, Google does not offer this option yet.
