How artificial intelligence is helping to restore works of art

Art restoration is a discipline that combines passion, patience and resourcefulness. And it is benefitting more and more from the latest technological advances, especially in terms of artificial intelligence. The start-up Oxia Palus is training one such AI to accurately reconstruct paintings that have been lost for years.

This year, visitors to the Focus Art Fair were treated to a special kind of exhibition at the MORF Gallery booth. It was a selection of paintings by Vincent van Gogh and Leonardo da Vinci that the general public has never previously been able to see because they are covered over by other works by the two artists.

One of them, “Two Wrestlers,” represents two men, shirtless, in the midst of a wrestling match. Vincent van Gogh probably painted it in 1886, while at the art academy in Antwerp. He alludes to it in a letter to his brother, Theo.

“This week I painted a large thing with two nude torsos — two wrestlers […] and I really liked doing that,” it reads. However, the Dutch painter covered it over with a still life, titled “Still life with meadow flowers and roses,” currently kept in the Kröller-Müller Museum in the Netherlands.

ALSO READ: This American novelist’s books are sold out thanks to BookTok

The Oxia Palus start-up recreated this underpainting by teaching artificial intelligence how to reproduce Van Gogh’s style. 

From Modigliani to Leonardo da Vinci

The researchers trained a neural network on other works by Van Gogh. They thus taught it how to reconstruct “Two Wrestlers” with precision from an X-ray-style image. This gave rise to a result that is “very convincing — by far the best guess we can get with current technology,” according to Anthony Bourached, one of the co-founders of Oxia Palus, speaking to Artnet News. 

This artificial intelligence has also managed to accurately reconstruct “Standing Female Nude” by Vincent van Gogh, “Madonna” by Leonardo da Vinci, “Beatrice Hastings” by Amedeo Modigliani and “Parc del Laberint d’Horta” by Santiago Rusiñol. 

These high-tech replicas have been dubbed “NeoMasters.” They have the particularity of offering a precise idea of the original color and texture of the long-lost paintings. Oxia Palus’ aim is to continue to put technology at the service of art restoration. “With potentially thousands of works of art hidden dormant beneath existing paintings, destroyed and missing, the journey in resurrecting the world’s lost art has only just begun,” reads the start-up’s website.

DALL-E 2, the robot artist

Proof of this ambition can be seen in the firm’s latest project, “TextMasters.” This resembles “NeoMasters,” with the key difference being that the AI reconstructs lost paintings from textual descriptions, not from X-ray images — a feat made possible by DALL-E 2, the innovation developed by the American company, OpenAI.

This text-to-image artificial intelligence software uses language comprehension models, as well as learning without human supervision on very large amounts of data, to generate credible artificial images. Oxia Palus researchers have already used DALL-E 2 to reproduce paintings by Diego Velázquez, Giotto, Leonardo da Vinci, Johannes Vermeer, Eugène Delacroix and Sandro Botticelli, among others.

Some of them are currently being offered for sale as NFTs on the KnownOrigin platform, in order to attract the interest of crypto-collectors. “We hope that by starting with some of the most notable lost pieces in history, we will be able to generate wider public interest in restoring our collective cultural heritage,” said Oxia Palus.