The capabilities of our artificial intelligences are continually improving. A site now offers to create photos of beaches that do not really exist. Ideal for making your friends believe that you are spending your holidays with your feet in the water.
If you want to show off on Instagram by posting pictures of heavenly places, why not use pictures of beaches that don’t exist? The site Thisbeachdoesnotexist.com (which literally translates to “this beach does not exist”) allows you, thanks to the prowess of artificial intelligence, to generate photos of completely fictitious coastlines.
The limits of AI
The tool is quite similar in its operation to the one that allows you to create people who don’t exist (or even cats that don’t exist). It relies on generative antagonistic networks (or GANs for generative adversarial network in English). This type of neural network is very much in fashion right now. It is also a GAN that made it possible to create this dreamlike version of GTA V.
By training artificial intelligence with some 20,000 beach photos, Czech researcher Vojtěch Semeckýn has succeeded in creating a tool capable of generating landscapes in the style of Mediterranean or Seychellois beaches. On the site, it is possible to create endless photos that smell of hot sand and holidays.
Not all images are perfect. We can sometimes see small failures, with poorly controlled blurs or palm trees which part of the trunk is missing. According to the developer, these errors are due to the fact that neural networks today are particularly good at understanding and creating ” images with a separate object in the center “, and ” the landscapes are quite the opposite. “In beach photos, errors are particularly visible in the finer details, such as on the foliage of trees for example.
What exactly is a GAN?
To successfully create this tool, Vojtěch Semeckýn was inspired by Nvidia’s research in the field and put a machine (from Google) to work for a month and a half. The operation of a GAN is quite simple. A first neural network (called a “generator”) will create range images, while a second (called a “discriminator”) will compare this output to real images to detect faults and improve the work of the generator.
If analyzing 20,000 beach photos for a month and a half already seems enormous, in reality it is not so important. The richer the photo database and the longer the generator and discriminator work hand in hand, the better the results. The principle of a GAN is that it automatically learns from its mistakes.
Still, the photos are compelling enough to make your Instagram friends drool with envy, making them think you’re lounging on a beach in the Maldives, while they suffocate in the summer heat.
