Picasso Painted Over This Portrait of a Sitting Woman. Now an AI Has Reconstructed the Lost Work of Art

It is not often that artists surprise us long after they die, but there are some notable exceptions. Think: Leonardo da Vinci, Rembrandt, Van Gogh and Picasso.

Underneath Pablo Picasso's famous Blue Period piece The Old Guitarist (1903-04), there is the faint figure of a woman. She was first detected in 1998, when conservators at the Art Institute of Chicago used X-rays and infrared light to peek behind the top layer of paint. She sits with her left arm reaching toward the viewer and her right arm tucked onto her lap—and curiously, she matches a sketch Picasso sent to a colleague around the same time.

Until now, all that could be detected of the woman was the faintest trace of her outline. Any extra detail—like color and style—was lost. However, by using a neural network trained to distinguish the style of one artist from another or (in the case of artists like Picasso) of one period from another, researchers at University College London have given her a new life, even if it is not necessarily true to the Picasso original.

A neural network has helped reconstruct a lost work from Picasso's Blue Period, circa 1903. Credit: Anthony Bourached and George Cann

The basis for the neural network was a technique developed in 2015 by Leon Gatys at the University of Tübingen, Germany, called neural style transfer. The machine vision technique can identify the style of a painting and transform the style of a painting to match another. It can transform a Michelangelo into the style of Paul Cézanne, for example. The technology has even been applied to videos—so if you want to watch clips of Ice Age in the style of Van Gogh's The Starry Night, you can.

Researchers Anthony Bourached and George Cann published their paper, titled "Raiders of the Lost Art," on the arXiv preprint server, and it was featured in a recent article in MIT Technology Review. They took the X-ray images of the woman's outlines and plugged them into a neural network built to transform images into the style of La Vie (1903), another masterpiece of Picasso's Blue Period.

While the result cannot exactly recreate the painting Picasso created—and subsequently painted over—over a century ago, it is at least typical of the style he was adopting at that period.

The same technology was also applied to Picasso's The Crouching Beggar, which was painted over a landscape thought to have been made by a contemporary Spanish painter, Santiago Rusiñol, using Rusiñol's Parc del Laberint d'Horta.

Bourached and Cann wrote that they hope the results encourage similar projects on other works of lost art, as well as challenge the commonly held "skepticism that an algorithm can be creative, or innovate."

"We argue that the use of machine learning as an artistic tool can ultimately broaden creative insight and widen the landscape of inspirational ingenuity by human-AI collaboration," wrote Bourached and Cann.

The Old Guitarist and The Crouching Beggar are not the only works of Picasso that mask hidden paintings. Indeed, artists frequently paint over earlier pieces, particularly when a lack of money might make fresh canvas hard to come by. In 2014, researchers found a portrait of a man resting his hand on his head underneath Picasso's The Blue Room (1901).