Google's Text-to-Image AI Can Create Any Imaginable Image
Google's Text-to-Image Artificial intelligence can transform descriptions such as "a couple of robots in a fine dinner with the Eiffel Tower in the background" into images.
Artificial intelligence appears to have reached a new milestone, this time at the hands of Google. A program called []Imagen promises to turn any descriptive text into an image. Any text, even the craziest ones you can think of.
The tool's website includes some extremely intricate, if not bizarre, examples:
- A magnificent oil painting depicting a raccoon queen in a red French royal gown. The painting is displayed on a well-decorated wall.
- In Times Square, a photo showed a Corgi dog riding a bike. He's sporting a beach hat and sunglasses.
- On a farm, there's a huge snake. Corn is used to making the snake.
- In every case, the end result is a stunning image. They appear to have been created in a drawing, 3D modeling, or photo-editing application.
The page brings a small demo, much more limited than artificial intelligence promises. In it, you have options for image style, animal, clothes, accessories, activities, and scenarios. Just match and receive your image.
The results are almost always great — the pictures of panda and raccoon wearing sunglasses didn't work, but otherwise, the images fit the description perfectly.
According to Google, its artificial intelligence beat OpenAI's DALL-E in a benchmark the company created itself. The test consists of having programs create 200 images from predefined descriptions and then putting them under human evaluation.
Text-to-image templates like these would allow people to create illustrations and montages without having to know how to operate programs like Photoshop, for example. But not now.
Artificial intelligence carries risks
Google has not released the tool for testing. Therefore, it is not possible to say if any description will turn out to be such a good image or if the company only selected the best results to show – assuming that there was no manipulation, obviously.
But there's a good reason for that. If Imagen is as powerful as it promises, it can be used for fake news, bullying, and harassment. Google also points out that the algorithm learns from images from the internet, so it can code for social biases like racism, sexism, or toxic behavior.
[]DALL-E, Imagen's “competitor”, is in beta and available only to selected users. It filters text entries to prevent the template from being used to create racist, violent, or pornographic images.
Google says Imagen is not suitable for public use at this time.