Restoration process of the painting. On the left is the original with damaged areas, and the middle shows various types of damage maps. The green line indicates cracks in the canvas, the thin red line denotes areas where the paint has cracked, blue represents areas without paint, and pink shows small defects that resemble scratches. On the right is the final restored painting, which is a digital restoration printed on film and attached to the original./Courtesy of Alex Kachkine

In 2012, an incident occurred in Spain when a mural of Jesus in a 19th-century church was ruined during restoration efforts. An 80-year-old woman from the town of Borja in the province of Zaragoza entered the church and repainted the mural of Jesus, turning it into a monkey-like image that was completely different from the original. It was considered the worst restoration in history, but fortunately, a painting of the same image by the original artist was discovered later, allowing for a proper restoration.

Restoring ancient paintings can only be done by highly skilled professionals. For centuries, preservation experts have restored paintings by identifying damaged areas and mixing the colors used at the time to refill them according to the original tones. Because this is done by hand, restoring one painting can take anywhere from several weeks to over ten years, depending on its condition.

Thanks to artificial intelligence (AI), the restoration of paintings that used to take years can now be shortened to just a few hours. AI learns painting information as digital data and restores damaged sections in the painter's style. Notably, AI has succeeded in restoring actual paintings to their former physical condition, going beyond digital restoration.

◇Combining digital restoration with actual paintings using AI

Alex Kachkine, a researcher in the doctoral program in mechanical engineering at the Massachusetts Institute of Technology (MIT), noted in an international journal, "Nature," on the 12th that AI restored damaged parts of a 15th-century painting in just three and a half hours.

This restoration did not involve any alteration to the actual painting. After AI digitally restored the damaged sections, the information was transferred to a film mask, which was then applied to the painting. The mask can be removed at any time.

Some argue that the damage to the painting is part of its history and object to restoring it by repainting. If it later becomes clear that the restored painting differs from the original, it would necessitate further damage and restoration. The film attachment method of restoration has been assessed as free from controversies surrounding original damage.

Kachkine said, "If the digital file of the mask is stored, future preservation experts can refer to it. Someone dealing with this work a century from now will be able to know what work has been done on the painting," adding that it would also be easier to correct any limitations in the technology used during restoration.

그림 복원 과정. 원본을 스캔해 각종 방식으로 손상된 부분을 찾아낸다. AI가 이 부분을 복원한 그림 파일을 만들고, 얇은 필름에 인쇄한다. 필름을 원본에 붙이면 복원이 끝난다./Alex Kachkine

◇Actual application in restoring 15th-century oil paintings

Kachkine applied this method to a severely damaged 15th-century oil painting by an anonymous artist. The painting depicts the Magi worshiping the infant Jesus. First, the paint added during the previous restoration was removed, and the original was scanned. AI identified 5,612 areas in need of restoration. It primarily sought out areas where colors had faded or paint had cracked.

AI learned to fill in the gaps by studying the digital information of the painting. For example, if a pattern in a curtain was repeated, the gaps would also be filled similarly to the front and back, and where some paint was missing from a face, it was recreated using surrounding colors. The face of the infant Jesus had completely disappeared. AI sourced a face from another painting depicting a similar scene and modified it to fit the style.

AI filled in the damaged parts with 57,314 different colors. Next, the restored segments were printed onto a film mask. A film painted white, representing the original canvas, was placed underneath those printed sections. This was necessary to create the effect of paint applied to the actual canvas. The two films were attached to the painting to complete the restoration.

The entire restoration process took three and a half hours from start to finish. Kachkine estimates that it was about 66 times faster than traditional restoration methods. While AI has been used for painting restoration in the past, this is the first time a digital restoration print has been applied physically to the original in the form of a film.

The process of AI restoring the damaged areas (red square box) in a 15th-century oil painting. It restored the empty spots by matching the surrounding paintings or referencing other paintings depicting similar scenes./Courtesy of Nature

This research began purely out of personal curiosity. Kachkine revealed that he has loved painting since he was a child and also did some restoration as a hobby. While traveling to MIT for his doctoral studies in 2021, he suddenly realized that there were more paintings in storage than exhibited works in galleries. He thought that if a faster restoration method could be developed, more people could actually see more paintings, prompting his research.

In a commentary published alongside Kachkine's work in Nature, Hartmut Kutzke, a professor at the Museum of Cultural History at the University of Oslo in Norway, stated, "While damaged paintings have been digitally restored in the past, there has been a dilemma of needing to view the original alongside the digital restoration displayed on screen. Kachkine resolved this issue by physically applying digital restoration to the painting."

He added that using this method to quickly restore damaged paintings allows them to be pulled from storage and publicly displayed, thus broadening public access to art. However, he pointed out that one must also consider the risk of further damage occurring due to chemical reactions taking place in the air layer trapped between the painting and the film.

Based on the sketch hidden beneath 'The Blind Man's Meal' painted by Picasso in 1903, the artificial intelligence created 'Lonely Curled Nude'./Courtesy of Oksia Palus

◇AI completes Picasso's unfinished nude work

AI not only restores damaged sections of paintings but can also create entirely new portions that were not drawn at all, completing the unfinished artwork according to the style of the artist.

Oxia Palus, a British art restoration company, announced in 2021 that it discovered a nude painting of a woman hidden beneath Picasso's work "The Blind Man's Meal" and brought its colors to life using AI.

This work is from Picasso's Blue Period (1901-1904). At that time, Picasso, who was an unknown artist, expressed the loneliness and misery of a life at the bottom through deep blue hues, fearing that he might go blind due to malnutrition.

Oxia Palus was co-founded by two researchers pursuing Ph.D. programs in AI at University College London (UCL). The two used X-ray fluorescence technology to identify the outlines of the nude painting beneath The Blind Man's Meal. They then trained AI to learn Picasso's Blue Period style.

Ultimately, AI painted oil colors onto a sketch of the female nude in the manner that Blue Period Picasso would have done. A 3D printer applied paint to the actual canvas according to the height information provided by AI. The painting was named "The Lonesome Crouching Nude."

References

Nature (2025), DOI: [https://doi.org/10.1038/s41586-025-09045-4](https://doi.org/10.1038/s41586-025-09045-4)