While the field of AI-generated imagery is still relatively new and often thrilling (or terrifying), much of what's coming out of research labs now is incremental improvements on already-existing technology. It feels like every week some institution puts out a new, slightly improved algorithm or pushes a whitepaper to arXiv and it's another dancing skeleton or Mona Lisa moving her lips all weird.
And then, sometimes, you get something that machine learning engineers and fans of the Animorphs books can all get excited about. That's what educator and AI researcher Xander Steenbrugge has done, with a series of AI models that morph humans into animals and back again.
Like many image style transfer creations, this one was made using generative adversarial networks, or GANs, which "learn" from a training dataset and then try to turn a target image into the source you feed it. For example, deepfakes are made using GANs.
To create what he calls "humanimals" but what can all clearly see are animorphs, Steenbrugge said he used a dataset of 15,000 HD animal faces that was published with a new generative model, StarGAN v2. He used that dataset combined with a set of human faces as the training data, and ran it through a different model, StyleGAN v2.
The result is this inter-species AI dream, where a woman becomes an eerie cat-person that becomes a man-leopard with aviator glasses. It's all very Zootopia.
"Most academic papers publish a certain method with an intended purpose ‘A’ in mind," Steenbrugge told Motherboard. "But because machine learning methods are so flexible, it often turns out that you can also do B, C and D with the same model, by just running it on different data and that is often where all the fun is at."
This experiment is part of Steenbrugge's larger project, Neural Synesthesia, where he creates immersive audiovisual experiences using algorithms and machine learning models. "I believe we are witnessing the start of a new era in digital media where creativity is becoming an interactive process between the human and the machine," he said. "I do not fully create these works, I co-create them with the AI models that I bring to life."