Using AI to create artwork may seem paradoxical, but the creative work of computers is increasingly being seen as of legitimate value. In a recent auction at Christie’s, the world’s largest and most famous auction house, an AI-assisted artwork sold for $432,500, leading critics to speculate wildly about the future of AI artwork, and whether this can be considered artwork at all.
Whichever side you are on, artistic AI is gaining traction, with artists adopting it in many different ways just as they would use any other tool to create original work or recreate the world’s most valuable pieces. With priceless artworks subject to collateral damage or vandalism, and prominent artists recently making bold statements about our attitude towards ‘priceless’ art, could computer-generated paintings change how we think about art and its value?
The RobotArt competition, founded in 2016 by Stanford graduate Andrew Conru, attracts a mixture of artists, mechanics and computer scientists interested in advancing the field of robotic art. This year’s winner CloudPainter (built and programmed by Pindar Van Arman) uses deep learning, cameras and a selection of 3D printed brush heads to achieve an increasing amount of autonomy in the paintings that CloudPainter produces – but as Van Arman stated in a Vice & HBO documentary last year, ‘my robots… make the same creative decisions I do,’ which seems to imply that he is still the ‘artist’ despite allowing the program to make decisions based on his instructions.
The third-placed entry by CMIT ReArt from Kasestart University in Thailand takes a different approach, and recreates an artist’s brushstrokes precisely including the force applied and position of the brush, to create an exact copy of the painting stroke by stroke. This method illustrates where the current approaches to making AI artwork diverge: some artists seek to make AI that is independently creative, whereas others try to recreate human abilities in form and function. While ‘original’ AI artwork is creating a lot of noise and debate in the artistic community, AI reproduction is becoming far more accurate and could solve problems that have plagued the art world since the dawn of time.
Recreating the classics
Reproductions have long been an important part of the art world, adding value to the original and allowing greater democratization of an artist’s work – in fact, print editions and reproductions have recently become the fastest growing segment of the art market. Faithful replicas do more than just spread an artist’s work. Replicas that are so close to the original as to be deemed a facsimile can be used to replace and preserve the original artwork, or revive classic paintings lost to fire, theft, or the mists of time. Creating these facsimiles historically has been difficult for a number of reasons, but new techniques have surfaced that address these issues while seeking to make the process of art reproduction easier and more accurate.
A team of researchers from MIT CSAIL has come up with a new method of reproducing paintings using deep learning and 3D printing. The new system ‘RePaint’, uses a mixture of two methods to accurately recreate color: a new technique called ‘color contoning’ which uses 10 thin layers of transparent ink, and ‘halftoning’ which is a much older technique that uses lots of ink dots to create a continuous color, similar to the classic painting technique of pointillism. This mixture of techniques solves three key problems in 2D printed reproductions: a limited color spectrum (of only cyan, magenta, yellow and black (CYMK)), a strict ‘total ink limit’ which if exceeded ‘results in deteriorated image quality, ink blotting, or mechanical malfunction’, and the fact that 2D printers can create a good reproduction under one lighting condition which is not accurate under another (a problem known as ‘metamerism’).
Using color contoning and halftoning, the 3D printer blends the mixture of ten different inks to create a range of ‘primary’ colors that match the target painting, and then places tiny dots of ink onto a canvas to reproduce the range of mixed colors, or halftones, that appear in the painting. Selecting which transparent inks to use in the ‘stack’ is the work of a deep learning program trained on images of original paintings under different lighting conditions. In this way, the program learns the ‘true’ spectral color of the painting, rather than picking up on light distortions in a 2D image, and can predict which range of inks will produce the most faithful reproduction under any lighting conditions.
Although RePaint still has a lot of improvements to make before it is commercially viable (certain colors like cobalt blue could not be reproduced by their range of inks), Changil Kim, one of the paper’s authors, says the ‘system works under any lighting condition and shows a far greater color reproduction capability than almost any other previous work.’ RePaint’s reproductions are also currently limited to the size of a business card due to the cost of 3D printing, but the team is hopeful that commercial printing can alleviate this expense once the project has been fully developed. This, however, may counteract some of the more altruistic consequences of creating artistic facsimiles, such as making culturally and historically important art more accessible to those who cannot experience the original.
Reproducing the goods
The innate creativity of AI may be a matter of contention, and the question remains whether art ‘created’ by AI should be viewed in the same way as a piece that took years for a human to conceive, but AI is unquestionably good at using input data to achieve a given output. When that data is color and the output is an indistinguishable copy of the ‘Mona Lisa’, does that decrease the value of Da Vinci’s original, or increase the value of the concept ‘Mona Lisa’ by allowing more people to appreciate its beauty?
Regardless of the perceived ‘talent’ of AI programs (and therefore the value that should be attributed to their work), creating faithful reproductions of famous art can help to preserve those paintings for posterity and avoid costly and time-consuming restoration (read: repeated damage to the ‘Mona Lisa’). RePaint may not be quite ready to tackle the world’s most precious artworks just yet, but by avoiding the race to bring about a true AI artist, the MIT CSAIL team may have created the world’s most talented reproductionist, without becoming bogged down in what ‘artistic talent’ should mean.