Boardwalk and theme park caricature artists might soon have some competition. Researchers from the City University of Hong Kong, Microsoft, and Tsinghua University demonstrated new neural networks capable of generating caricatures from both images and videos. Two generative adversarial networks (GANs), trained on thousands of hand-drawn sketches, were employed: The CariGeoGAN analyzes a face’s geometry and syncs it up with a caricature model, while the CariStyGAN optimizes with a caricature style. It turns out that even the exaggerated, surreal nature of caricatures, which must have at least some semblance of the original subject, only fooled 22.5 percent of people into thinking they were human-drawn. Still, this does seem like a cool functionality for a smartphone app of the future, and the researchers say it could also one day be used to identify people in caricatures.
sign up for our newsletter