Here is the GANimals tool.
Researchers working for Nvidia have developed a tool that can use artificial intelligence "imagine". like other animals they would look like even your dog. GANimal, because that is the name of the tool, transfers the facial expressions of the animal visible in the selected photo to the photos of other animals. Best of all, each of you can try this tool. Just go to this addressand then upload the photo of your pet to the system. If you don't have one, you can play using images found on the Internet.
As you will notice quickly, the tool does not always give perfect results. Animals that have been transferred to another animal's facial expressions often just look weird. However, this does not change the fact that the technology behind the tool is amazing.
"Most image translation networks that are based on GAN algorithms are trained to solve individual tasks – for example, translating horses into zebras." said Ming-Yu Liu, one of the Nvidia researchers. "In this case, we train the network so that it simultaneously solves many translation tasks, where each sentence translates a random source animal into a random target animal. Ultimately, the network learns to generalize to translate known animals into animals that it has not yet seen. "
It is worth recalling that GAN algorithms are Generative Opposing Networkswho learn with the help of so-called generators and discriminators. Usually, based on the information provided, the generator creates images that look like real images, and the discriminator, who receives both images generated and supplied from the resources of another neural network, must distinguish them from each other. The teaching process ends when the generator begins to create images so similar to real photos that the discriminator ceases to be able to pick up the differences. Based on such networks, Nvidia researchers developed an algorithm called FUNIT (Few-shot, UNsupervised Image-to-image Translation) that the GANimal tool uses.
The FUNIT algorithm is distinguished from the fact that you don't need thousands of photos to train it. In addition, once you have trained FUNIT, you only need one source photo and one photo of each target animal (none of which had to be seen before) to accomplish its task.
Although at first it is difficult to imagine what the applications of the described technology could be, in time some interesting ideas come to mind. Similar algorithms could be used, for example, in the film industry – where the abilities of trained dogs could be digitally transferred to animals that are more difficult to master.