Rembrandt has been dead for nearly 350 years, but that apparently hasn't stopped him from producing one more work of art. Well, sort of. Programmers and art experts have created a computer-generated painting that attempts to replicate the style and brilliance of one of the world's masters. Using data analysis, facial-recognition technology and 3D printing, the Next Rembrandt team produced a frame-worthy portrait, replete with brushstroke detail and layered textures.
Tu Thanh Ha explains how they did it.
It started as a branding campaign for a Dutch bank, but the results are eye-catching.
A team of software programmers and art experts in the Netherlands have joined forces to create a computer-generated painting in the style of the great 17th-century artist Rembrandt.
Unveiled this week in Amsterdam, the fruit of their efforts – a three-quarter-view portrait of a man with a fallen ruff collar – is not just a digital image but was printed in 3-D, with brushstrokes and canvas texture like a real oil painting.
"It's a virtual re-enactment of what you see in a Rembrandt painting," said Joris Dik, a professor in materials in arts and archeology at Delft University of Technology.
Prof. Dik, who acted as an adviser on the project, had previously made a 3-D reproduction of a genuine Rembrandt.
The latest project, dubbed The Next Rembrandt, began as a brief from the Dutch bank ING to the Amsterdam branch of the J. Walter Thompson ad agency.
ING had long been an arts sponsor in the Netherlands and sees itself as an innovator in financial services. "They asked us to create a brand image campaign for them, to show their innovative spirit in their sponsorship of culture and arts," said Jesse Houwling, digital director at J. Walter Thompson.
Mr. Houwling said his ad agency was inspired by a British medical artist who used ancient skulls to create a 3-D model of what could have been the face of Jesus.
The agency decided then to focus on Rembrandt, an iconic artist for the Dutch. The agency's technology director, Emmanuel Flores, partnered with a trio of programmers, Rembrandt experts and developers at Microsoft. It took 18 months to produce the portrait.
While not revolutionary, The Next Rembrandt project's results are impressive, said Craig Kaplan, a computer-graphics professor at the University of Waterloo.
Because it is not clear how much human intervention was required to fine-tune the final results, "my instincts tell me that this work does not represent a fundamental breakthrough. … However, it is a highly effective application of known techniques in the pursuit of a specific goal," Mr. Kaplan said in an e-mail interview.
Mr. Flores said the methods they developed could be a stepping stone toward creating algorithms to restore or authenticate paintings.
Painting by numbers 1011100110
To develop algorithms that could create a Rembrandt-like work, the team first had to gather data about the master's works, scanning and analyzing 346 paintings attributed to him.
While Rembrandt is famous for some group paintings such as The Night Watch or The Anatomy Lesson of Dr. Nicolaes Tulp, commissioned portraits were more common. Focusing on them would maximize the data available. The developers thus settled on a typical painting as being a portrait of a white man with facial hair, facing to the right (women's portraits faced to the left).
They then extracted features from similar portraits to use as building elements for their virtual painting.
The first big challenge, Mr. Flores said, was generating a set of Rembrandt-like facial features that was not generic-looking.
If the computer had merely amalgamated, then averaged the measurements for all the noses, eyes and mouths of the original portraits, "you will have a blurry image," Mr. Flores said.
They had to design algorithms that mapped the colours and locations of facial pixels, so they could identify those that were recurrent and more relevant.
"So in the end, what we came out with was what, from a statistical perspective, represented the typical set of features used by Rembrandt."
At one point, the team went through a tryout phase where they printed an array of partial faces with varying colours and shapes.
After the subject was fully formed, the next challenge was to determine the overall lighting, the background and colour gradients of the virtual painting, which was still a two-dimensional image at the time, Mr. Flores said.
Rembrandt was famous for his skillful use of light and dark, to the point that his name is used to designate a classic setup in portrait photography, where one side is illuminated, along with a small inverted triangle of light on the subject's shadowed cheek, opposite the light source.
The final hurdle was creating a 3-D reproduction, with its own topography. They turned to Prof. Dik and Delft University, who had previously made a high-resolution scan of a Rembrandt work, Saul and David, then made a 3-D reproduction of it last year.
The Next Rembrandt simulation came out of a special printer, with 13 layers of ink layered on top of each other.
"They not only created the colours of a virtual Rembrandt but they also created the texture of the surface. It's not a flat surface, you see the painting's cracks. You see to some extent the texture of the canvas. You see the brushstrokes," said Prof. Dik.