Researchers from Google Research remain at the forefront of artificial intelligence technologies. Their latest creation, Imagen, is able to produce any photo from a simple written text. It is therefore sufficient to fill in the input field with “photo of a Corgi dog riding a motorcycle” so that Imagen produces in a few seconds the photo corresponding to this description (photo of course created from scratch).
Some fake photos produced by Imagen, with under each of them, the descriptive text on which the AI relied to produce the final image
The results are often staggering – the AI really can produce any type of image – but Google engineers say the software still has severe limitations. Thus, the AI struggles to create fake photos whose content is too far from the huge image bank it has (and from which it composes its own images). A dog riding a motorcycle is fine, but a horse riding a man doesn’t work.
On the left, the images produced by Imagen, on the right real photos corresponding to the expected result
Sometimes, it is the “associations” that do not correspond to the expected results. If Imagen is asked to produce an image of a panda made from coffee cream, the AI generally produces a “photo” of a panda preparing a cappuccino! Finally, Google researchers still fear the misuse that could be made of these technologies (racist, pornographic content, etc.), so that Imagen is not ready to be available to the general public. These considerations are certainly commendable, but by dint of caution, there is no risk of one day witnessing the expected revolution in automation and the management by AI of the most repetitive tasks, with the risk that only large groups benefit from it…