Using a new algorithm from Nvidia, you can face-swap your pet’s sweet mug onto other breeds to create monstrous versions of themselves.
PetSwap is a project that uses Few-shot UNsupervised Image-to-image Translation, or FUNIT—an image translation AI algorithm. FUNIT uses a framework built on generative adversarial networks, or GANs, the same framework that Nvidia has used to generate people who don’t exist and train autonomous vehicles on driving conditions that don’t exist. It’s also similar to how deepfakes are made, which were first inspired by Nvidia’s GANs algorithms.
“Humans are remarkably good at generalization,” the researchers wrote in their paper about PetSwap, which is available on the arXiv preprint server. “When given a picture of a previously unseen exotic animal, say, we can form a vivid mental picture of the same animal in a different pose, especially when we have encountered (images of) similar but different animals in that pose before.”
For example, our brains can easily imagine what a standing dog might look like when it’s lying down, based on what we’ve seen other animals do. But machine learning algorithms can’t. With FUNIT, researchers are trying to fill that imagination gap. Having an algorithm assume what something might look like based on other examples would allow it to use a smaller dataset of training images.
Nvidia made a demo available for the public to test out its algorithm, so I chose to faceswap my roommate’s cat Cookies, a creature of unknown breed and gender. By drawing a tight rectangle around Cookie’s cross-eyed face and hitting “translate,” the algorithm gave me several mutant animals in an attempt to match the cat face to new species’ bodies. The results are unsettling.
Having failed with Cookie’s perfect face, of course I had to try it on my own. The algorithm isn’t built to translate human faces, but it was worth a shot.
It looks like I’m midway through an Animorphs transformation. I’m a pretty cute mongoose, though.
Having an algorithm that could generate new images based on previous knowledge, rather than huge amounts of data, would make training a faster, less onerous task for researchers developing machine vision—such as self-driving cars, helper robots, other autonomous tech. But for now, the FUNIT algorithm is still just a fun way to deepfake your pet.