Microsoft researchers use AI to generate caricatures from photographs

A group of researchers have created an AI based app that can take ordinary photographs of people and turn them into hilarious caricatures (via The Next Web). The researchers, Stanford graduate Kaidi Cao, Microsoft’s Lu Yuan, and Jing Liao from the City University of Hong Kong are expected to unveil the project next month at SIGGRAPH Asia in Tokyo.

The AI was trained from more than a thousand different photographs, and uses a pair neural network technologies know as generative adversarial networks, or GANs for short, to generate the caricature. The first one looks at the geometry of the face to create the base model, while the second one applies an artistic texture to the model to give it the hand-drawn look.

The model can create muliple generations using the original and two references.

While the researchers admit the AI isn’t perfect, but almost a quarter of the generated images can be mistaken for actual human-created caricatures. The project is much like the one Microsoft Research brought us in January, which turns photographs into what appear to be painted portraits. However the new project give the art a humorously exaggerated effect.

What do you think about the project and how it can be improved? Feel free to share your thoughts below.

Further reading: Artificial Intelligence, Microsoft, Neural Networks