Google’s text-to-image AI is incredibly impressive… scary impressive
We’ve already seen demos of a "text-to-image" engine scattered across the Internet — an algorithm called "DALL·E" that’s able to accurately understand a string of words and digitally paint whatever was requested.
Well, little did we know — the Google Brain team, which works specifically on AI, was also developing its own text-to-image engine. It is called Imagen and its work is demoed on its official website. We have to say, the results look nothing short of impressive. Both awe-inspiring and scary impressive.
See, Imagen takes things a few steps further. When you are inputting your request, you can specifically ask the engine to generate a painting or simulate an actual photo. Yes, you can request a photo of a golden retriever wearing a bonnet and the AI will do its absolute best to manifest one out of thin air.
In order to gauge the quality of the Imagen's work, the Google team generated images from a selection of AIs and presented them to a group of people to assess. Across the board, according to the report, people liked Imagen’s results best.
Some of Imagen’s work
It’s kind of crazy to consider the implications of such advancements in AI. We already see apps that can generate actual music, based on deep learning algorithms. Now, we see AI creating images that can easily be used for logos, flyers, branding, et cetera. The one thing that we thought will always remain out of bounds for machines — art — seems to not be as exclusive to the human mind as we believed.
This technology, same as Dall·E, isn’t going to be made available to the public any time soon. Google is worried that once such a tool is out in the wild, there will be an avalanche of realistic-looking offensive imagery all over social media. Plus, with such advanced photorealism, the fear of deepfakes is real.
Supposedly, researchers are considering some sort of framework or limitations in order to make text-to-image engines less exploitable by trolls. But, until a solid solution is found, we are pretty sure no researcher out there would want to throw their AI baby to the Internet sharks.
Things that are NOT allowed: