Caricatures by Artificial Intelligence

These caricatures were created by artificial intelligence, not by humans.

The system uses Generative Adversarial Networks (GAN), and doesn't require any input from humans to create the caricatures. Scientists developed this AI system to use a two-stage process, a combination of style transfer and geometric exaggeration of facial metrics.

The artistic styles are developed by training the network with large data sets of human-created caricatures from Toonpool and Pinterest. Ideally you'd want to give the computer examples of direct comparisons between photos of people and artistic interpretations from those specific photos, but there's just not a large enough set of those to use for training the computer.

Then they trained the system to distort the artistically-styled image to emphasize the factors that make the face differ from an average face. The scientists explain: "Shape exaggeration is not a distortion, which is complete denial of truth. The exaggerated shape should maintain the relative geometric location of facial components, and only emphasize the subject’s features, distinct from others. The final appearance should be faithful to visual styles of caricatures, and keep the identity with the input face, as addressed in other face generators." 

The system can also be used for animation. These are stills from a video where each frame is created by the computer.

Most of the reactions I've seen to this fall into two camps: "It's coming to take my job" or "It's not that good." To the first I'd say it depends on what your job is, but it might force many of us to change our jobs somewhat if we want to stay in business. For some artists who want to use these tools, more power to you, and these methods may lead to new art expressions that we've never seen before. 

To the second objection, I'd say that this is just a first baby step. How far the neural network strategy will take us is anyone's guess. Some say GANs are superficially impressive but that the method has hit a wall, and it can't begin to rival the complexity and sophistication of the human brain. 

As blog reader Eugene Arenhaus puts it: "It tells more about our preconceptions about what our brain can do, than about AI. Turns out certain functions require far less brain than previously thought - and not unique to humans, as a less snobbish look at animal intelligence had been showing in the past 15 years. But there is a lot of complexity in human cognition that these primitive NNs can achieve, although the people driving hype try hard to pretend otherwise and make wild promises."

For my part, I don't regard this as a direct threat to my livelihood, but more of a reminder that I'll have to evolve my art as these tools get more powerful—similar to the way artists had to evolve with the advent of photography. I also am fascinated to learn how these AI systems work, because those insights can help me understand better my own human thought process when I make a picture.

-----

YouTube video: Can an AI Learn to Draw Caricatures?
Scientific Website: CariGANs: Unpaired Photo-to-Caricature Translation

PDF of Science Paper