Artnet by Jo Lawson-Tancred June 12, 2025
The conservation of old paintings is the latest slow and painstaking process that might soon be sped up by artificial intelligence. A mechanical engineering student at MIT has just debuted a cutting-edge A.I. technique that could allow an aged or damaged painting to be restored in just a matter of hours.
When it comes to conserving art, the stakes are high. Everyone has at some time marveled at the horrors of a botched job and these interventions are not always reversible. In recent decades, however, conservators have developed highly sophisticated methods relying on scientific advances like x-ray imaging and pigment analysis.
These methods help conservators to better diagnose the problems plaguing historical masterpieces, which often show their age through discoloration, a cracked or flaking surface, and the build up of dirt. Most of the work, however, like cleaning the surface or repainting lost patches of the original composition, is still done by hand.
Scans of a 15th-century painting during various stages in its restoration. On the left is the damaged piece, with the middle panel showing a map of the different kinds of damage present, and at right is the restored painting with the applied laminate mask. Image courtesy Alex Kachkine.
For this reason, the restoration process can take months and requires significant funding. Only the most important works are able to get top-notch treatment, but this could soon change thanks to a new A.I. technique developed by MIT student Alex Kachkine. Their idea for the product arose from their own passion for collecting historical artworks. Their budget could only stretch to damaged pieces so they began restoring art by traditional means as a hobby. One day, they decided to see if they could apply their engineering know-how to invent a quicker fix.
“If we could just restore a painting digitally, and effect the results physically, that would resolve a lot of pain points and drawbacks of a conventional manual process,” Kachkine realized.
The technique works by taking a high-resolution scan of the artwork and using a pre-existing A.I. algorithm to identify cracks or patches where the original composition has been lost. These areas are then digitally restored, a process that has already been trialed by some other experts in the field. Until now, however, there has been no way to transfer these digital alterations onto the physical artwork.
Kachkine seems to have solved this dilemma with what they call a “digital mask,” created by printing the digital restoration onto polymer-based films using high-quality pigments. This mask can be overlaid onto the painting’s surface and sealed in place with a varnish. Crucially, if necessary, the mask can also be removed without a trace using conservators’ solvents.
A diagram showing an exploded view of the final reconstruction composition. Image courtesy Alex Kachkine.
According to a paper authored by Kachkine and published by Nature, 70 percent of paintings in museum collections are kept in storage and this is partly due to the exorbitant cost of restoring them for public view.
The engineer goes on to outline the process of using their new tool to restore a badly damaged 15th-century oil-on-panel painting by the Master of the Prado Adoration of the Magi. The A.I. identified a whopping 5,612 sections of the painting that were in need of repair. During the digital restoration, missing patches were color matched to their surroundings or, in the case of more complex patterns, the section was copied from elsewhere in the painting.
Kachkine wrote that the infill process, the application of the mask containing 57,314 colors to the painting, took just 3.5 hours to complete. They estimate this to be 66 times faster than the conventional approach of inpainting by hand.
“This approach grants greatly increased foresight and flexibility to conservators,” Kachkine’s paper concludes. “Enabling the restoration of countless damaged paintings deemed unworthy of high conservation budgets.”
As a dedicated art lover, Kachkine is aware of the ethical complexities of restoring art, whether by hand or A.I. The appropriate degree of intervention can come down to the judgement of the individual conservator. For example, while restoring the Master of the Prado painting, Kachkine replaced the missing head of an infant with an example take from another work by the same artist.
Details of masking results during restoration process. Image courtesy Alex Kachkine.
Not only are all changes made using Kachkine’s technique reversible, they come with a digital record of the changes made so that future conservators will have a full understanding of what treatments have been done. So far, Kachkine said, the reception to the technique has been “cautiously optimistic.” They are currently fundraising to continue its development, with the hope that it may one day be easily adopted by interested conservators.
Kachkine’s breakthrough may allow conservators to apply digital restorations to a physical artwork. This could advance pre-existing efforts by some art scholars and computer scientists to create digital reconstructions of lost art. One particularly high-profile example was the use of A.I. to retrieve the lost outer sections of Rembrandt’s The Night Watch, which was chopped down to its current size in 1715 in order to fit into Amsterdam’s Town Hall.
The project, spearheaded by the Rijksmuseum, trained A.I. to replicate the distinct style of Rembrandt and apply this to recreate the lost sections according to information about the original composition found in a copy of the original by artist Gerritt Lundens. These sections were then printed onto panels and installed around The Night Watch to recreate Rembrandt’s intended composition.
Conservators have also had some success using A.I. to decode x-ray images that are otherwise difficult to read. In the case of the Ghent Altarpiece, the paintings are on double-sided panels so it proved difficult for conservators to parse the x-ray images. In 2019, a newly developed algorithm allowed scientists to deconstruct the data to create two distinct images.