Everyone who keeps a pet at home loves them. However, would your favorite dog look even more cute if it was of a different breed? Thanks to a new tool from NVIDIA called GANimals, you can evaluate whether your favorite pet would look even cuter if it were a different animal.
Earlier this year, NVIDIA Research web users with its GauGAN tool, which turned rough sketches into near-photorealistic images. This tool required users to specify which parts of the image should be water, trees, mountains, and other landmarks by selecting the appropriate brush color, but GANimals works completely automatically. All you have to do is upload a photo of your pet and it will create a series of photorealistic images of other animals that will retain the pattern's 'expression'.
This week, in a paper presented at the International Machine Vision Conference in Seoul, Korea, the researchers described the algorithm they developed β . It stands for Few-shot, UNsupervised Image-to-image Translation. When using AI to transform the characteristics of a source image into a target image, the AI ββtypically needs to be trained on a large collection of target images with different light levels and camera angles in order to produce results that look realistic. But creating such a large database of images takes a lot of time and limits the capabilities of the neural network. If the AI ββis trained to turn chickens into turkeys, that's the only thing it'll be good at.
In comparison, the FUNIT algorithm can be trained using just a few images of the target animal, on which it is repeatedly practiced. Once the algorithm is sufficiently trained, it needs only one image of the source and target animals, which can be completely random and have never been processed or analyzed before.
Those who wish can try out GANanimals on , but so far the results are low resolution and not suitable for anything other than informational purposes or to satisfy curiosity. The researchers hope to eventually improve the capabilities of AI and the algorithm so that it will soon be possible to change people's faces without relying on huge databases of carefully selected images.
Source: 3dnews.ru