Machine Learning Generated Artwork Auctions Off for $ 432,500

Far from being the sole creation of AI, portrait of “Edmond de Belamy” was the result of months of work using machine learning

October 25, 2018 by Roberto Iriondo

Picture | Courtesy of artist group Obvious | [3]

A machine learning generated print sold for $432,500 at Christie’s auction house in New York on Thursday, October 25, 2018 — over 40 times higher than expected to sell.

The print is called “Edmond de Belamy” and its blurred-out features — are the artifacts of the generative adversarial networks[1] (GANs)— the machine learning algorithm used to generate it. The print is one of a series of 11, all ML-generated, and depicting the members of the fictional “Belamy family.” Belamy’s portrait is signed with the mathematical formula describing the algorithm that was used to generate it.

Generative Adversarial Networks | The machine learning algorithm that was used to generate the Belamy portrait [1]
Pseudocode of GAN training | Arxiv | [1]

Generative adversarial networks (GANs) are generative models created in 2014 by Ian J. Goodfellow, a machine learning researcher from Google Brain, who basically placed two algorithms in competitive mode with one another as to perform training.

How to train a GAN | A Short Introduction to Generative Adversarial Networks | Thalles Santos Silva [5]

Far from being the sole creation of artificial intelligence, the piece is truly the product of months of machine learning iterative work by three people living together in a Parisian flat — one of whom is a machine-learning PhD student — the group collectively call themselves Obvious [3].

Interactive Image Generation using GANs | Drawing basic strokes and let the model draw impressive pieces for you | Jun-Yan Zhu | [10]

The piece’s inclusion in the Christie’s auction, next to prints by Chuck Close and Jeff Koons, has been the cause of some consternation in the art world, but also among AI experts who take umbrage with the implication (by virtue of the signature on the piece) that an algorithm can be an artist all by itself — especially the relatively humdrum variety that was used to create the piece.

The algorithm is not the only element that went into creating these pieces — GANs do not have free will. They output complicated paintbrushes based on complicated mathematical-input parameters, and you can use these paintbrushes to achieve an effect that might be difficult to achieve otherwise.

The cajoled GAN algorithm adjusted by machine learning PhD student at Obvious [3]

This nuance was clearly lost in some of the reporting around the piece leading up to the auction — numerous headlines described the piece as being “created” by AI, as opposed to using machine learning or being generated with machine learning, which makes it obvious as to how poorly informed is the public is on the differences between AI and machine learning.

GANs started to receive attention from the artist community due to the potential of generating intriguing art pieces. Below please find a GAN generated art by Mario Klingemann [6], an artist who has used GANs numerous times to create interesting pieces.

“Freeda Beast — Bringing Things to an End” | A product of GANs training| Quasimodo | Mario Klingemann [6]

Hugo Caselles-Dupré, a machine learning PhD student and one-third of Obvious, stated [7], that he chalked this characterization up to “sensationalism” and “clickbait” in the media. The intent of the piece, he insisted, is to educate the public on the limits of artificial intelligence. Algorithms are a tool, Caselles-Dupré mentioned, not creative beings themselves.

“Today, it is not about algorithms that are replacing people,” Caselles-Dupré said [7]. “In the future, we may have to be careful about this, but today, they’re more like a tool. We really wanted to showcase a concrete example of what these AI tools can do.” Signing the piece with the algorithm’s mathematical formula was a “funny way,” he mentioned, as to communicate these ideas to a general audience.

“The Belamy Family” | Machine learning generated art | Obvious [3]

It is vague how well such messaging strategy has functioned. In a statement to Artnet [8] Richard Lloyd, Christie’s international head of prints and multiples said that the piece was selected for auction precisely because of how supposedly little human intervention went into creating it.

In order to make these models work, GANs must be fed a large amount of data, and use such image datasets to produce new results after an strenuous training period. However, just because they are capable to produce original outputs, GANs are not autonomous. Such end products are the result of a long process of carefully selecting input data, tweaking mathematical parameters, and then sifting through the results to find the best results of whatever it is that you are looking for.

The final iteration of the algorithm — the best it was ever going to get for Obvious’ purposes — spit out hundreds of images, Caselles-Dupré said, which had to be whittled down to just 11. “We carefully selected the images that we found the most interesting in this batch,” he said.

After several iterations of using GANs, Obvious [3] selected 11 art pieces which they carefully selected as they wanted novel results [7]. However, which is the creative party in this process: the algorithm that needs to be iterated and cajoled for months to turn out something half-interesting, or the artists searching for an aesthetic result and making all of the decisions to get there? While the output has been a collaboration between computer algorithms and the artists involved, the balance of creativity falls on the side of humans.

Such use of GANs will definitely not replace artists, however, it does bring a new perspective to the public and the art community on the use of machine learning algorithms as to generate novel pieces [11].

Density estimation using Real NVP | Laurent Dinh, Jascha Sohl-Dickstein, Samy Bengio[12]

Lastly, Obvious [3] mentions on their blog [9]:

“Will artificial intelligence be the artist of tomorrow?”
I would be tempted to answer:
“Is the camera the artist of today ?”


[1] Generative Adversarial Networks | Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio | Arxiv |

[2] Quasimondo | Mario Klingemann, Arist |

[3] Obvious | Paris-based collective of artist and machine learning researchers |

[4] Christie’s Auction Edmond de Belamy, from La Famille de Belamy | Obvious Group |

[5] A Short Introduction to Generative Adversarial Networks | Thalles Santos Silva |

[6] Mario Klingemann |

[7] An AI-Generated Artwork Just Sold for $ 432,500 at Christie’s | Motherboard |

[8] Has Artificial Intelligence Given us the Next Great Art Movement? Experts Say Slow Down, the “Field is in its Infancy” | Artnet News |

[9] A naive yet educated perspective on art and artificial intelligence | Obvious |

[10] Interactive Image Generation via Generative Adversarial Networks | Jun-Yan Zhu|

[11] Machine Learning for Creativity and Design | NIPS 2017 Workshop |

[12] Density Estimation Using Real NVP | Laurent Dinh, Jascha Sohl-Dickstein, Samy Bengio |

[13] Structured Generative Adversarial Networks | Zhijie Deng, Hao Zhang, Xiaodan Liang, Luona Yang, Shizhen Xu, Jun Zhu, Eric P. Xing | Tsinghua University, Carnegie Mellon University, Petuum Inc. | Nvidia Research Pioneer Winner |