How Do You Create a New Rembrandt?
A computer becomes the Dutch Master’s best pupil
With the help of a corpse, in 1632 Rembrandt van Rijn triggers a revolution. The 25-year-old draws a corpse being dissected by a doctor who is surrounded by students—and in so doing, breaks with centuries of convention in painting. Because the subjects depicted in his work are not shown standing next to one another in neutral poses as was customary at the time; instead, they are standing behind one another in lively poses. The most notable innovation: Faces reflect excitement and fascination. “Against the dark background their emotions look particularly strong, as if the painter had dipped his brush not in pigment, but rather in sunlight,” remarks psychologist Barbara Diggs. This play of colors is what makes the 346 known “authentic Rembrandts” unmistakable. This number, 346, was considered valid for the 346 years following the master’s death in 1669. Then, on April 5, 2016, the luminary’s 347th painting was unveiled in Amsterdam. It was created by a computer capable of simulating the artist. But what is “The Next Rembrandt” in actuality? And could an artificial intelligence really be creative?
To this day, aspiring painters learn to hone their craft by trying to copy the images made by the great masters. A computer can do the same—just much more diligently, with untiring persistence: Beginning with the knowledge basis of a young child, a computer network took several months to analyze how Rembrandt painted the details of a face—which proportions he chose, how he used colors. This process is called “deep learning.” Using a total of 168,263 image detail excerpts that were taken from all of Rembrandt’s paintings, a 20-person programming team eventually produced algorithms that deliver the same compositions as Rembrandt created at his easel. “We’ve distilled the painter’s artistic DNA out of his body of work,” says project director Bas Korsten. And building on these capabilities, the software can now even create its own motifs. The result: a fictitious 35-year- old with a beard, hat, and collar who looks as if Rembrandt drew him. The perfect forgery? This method does not permit that possibility: A chemical analysis of the color pigments would reveal the truth. The generated image corresponds with the abilities of Rembrandt around the year 1630, as his later works are still too difficult to attempt. But just as Rembrandt did not stop painting after that revolutionary anatomy lecture, so too will Microsoft continue this project. “Deep learning” has only just begun…
“WE HAVE USED TECHNOLOGY AND DATA AS REMBRANDT USED BRUSHES AND COLORS.” RON AUGUSTUS, MICROSOFT