AI image generation is nothing new, but Google’s latest research paper displays an advancement of a kind we haven’t seen since the first AI-driven image generators came about. In a nutshell, Imagen takes text and turns it into a realistic-looking image, driven completely by an AI that understands a large dictionary of words and what those words mean in a visual sense.
Google released its Imagen (opens in new tab) research paper alongside the explanation of the tool. Google says its own in-house developed benchmark, DrawBench, rates other image generation models based on human raters and show that the Google AI (opens in new tab)-powered Imagen produces superior results to those other models. The Imagen site showcases a number of different examples, a few of which we cherry-picked below.
Underneath the images above, you’ll see the text that was used to generate that image with Google Imagen. In many cases, the text is extremely descriptive and is used to create a very specific end result that looks surprisingly realistic. Many other AI-driven image generators that you’ll find online often create very abstract-looking imagery, as you might have seen on your favorite social media apps (opens in new tab).
But, while Google has plenty of examples and an entire research paper to show how well Imagen works, it isn’t making the technology publicly available just yet. In its explanation, Google sites societal concerns as the main reason for not letting users give it a shot just yet. Google says it believes that harmful, realistic imagery could be generated because the dataset used includes many uncurated words, many of which could be considered racist, derogatory, or otherwise harmful.