When lined up against Rembrandt’s work, it can be difficult to tell which portrait a machine created and which the Dutch painter created roughly 400 years ago.
The creators of the “new Rembrandt” used computers to 3D scan and analyze 346 Rembrandt paintings. Next, they used facial recognition software to identify the most common geometric patterns Rembrandt used. Once their computer system had learned this, it could replicate the style and create new facial features.
The process relies on a hot field in technology, deep learning, in which machines are fed massive amounts of data and are able to suss out patterns and then mimic them in new creations.
The team decided to create a portrait of a white male in his 30s with facial hair, dark clothes and a white collar, because those traits were so common in Rembrandt’s work. They generated the facial features individually and then assembled them into a face. The distance between features were based off calculations of what was typical in Rembrandt’s other works.
To create a look that mimicked the brushstrokes of Rembrandt, the researchers scanned the surface texture of Rembrandt originals to identify patterns in his texture. The portrait was then 3D printed with 13 layers of ink to create the appearance of Rembrandt’s brushstrokes.
The project took 18 months to complete, and the final product included more than 148 million pixels.
The work is reminiscent of a series of apps and online services that have sprung up in the past year, including DeepArt and Pikazo. These allow anyone to apply the style of a famous painter to an ordinary image.