As a first real project, we used Generative Adversarial Networks, a Machine Learning algorithm, to enable a computer to create portraits of the 18th century. We are Gauthier Vernier, Pierre Fautrel and Hugo Caselles-Dupré, three 25 years-old friends since childhood and we want you to hear our story, put in the right context.
Art created by Algorithms
A word about Generative Adversarial Networks (GANs)
Generative Adversarial Networks (GANs) are generative models created in 2014 by Ian Goodfellow, a researcher in Machine Learning. They put two algorithms in competition one with another to perform training.
A generator will create new images by mimicking characteristics of images from the training dataset, and try to fool a discriminator into thinking those images are “real”. The generator trains until no difference can be made by the discriminator.
Previous work on using GANs for Art
GANs started to receive attention from the art community, due to their potential to generate novel artworks. Until now, several projects produced inspiring images using GANs.
Notably, we have Mario Klingemann, an artist working with code, AI and data. He has worked with GANs numerous times, for instance you can see one of his artworks, that he created using a pix2pix model:
There are also research papers on using GANs for Art, and studying the creative potentiel of this algorithm. CAN (Creative Adversarial Networks) is a variant of GANs that are designed to produce “creative” artworks. In this study, being creative is defined by being able to produce an artwork that does not fit a classical art movement (at least, according to the discriminator).
In late 2017, a “Machine Learning for Creativity and Design” Workshop was organized at the most prestigious Machine Learning conference in the world: NIPS. Lots of submissions included work with GANs. The second edition of this workshop is to appear in December 2018.
Notably, we also have the work of Robbie Barrat on nudes, landscapes and portraits generation with GANs, from which we were inspired. Check out his work : https://github.com/robbiebarrat/art-DCGAN.
There are a lot more examples of Art with GANs, such as the work of Michael Tyka, Samim, Alex Champanard, and several research papers on the topic…
You should definitely check out their work if you want to discover more about Art made with AI tools.
Our first collection: the Belamy Family
Our first collection is a series of 11 realistics portraits generated by GANs. Out of these, we created a fictional family of eleven artworks, that we printed and framed.
Each portrait is a 70x70 cm artwork, with a golden wooden frame. It is almost similar to a portrait that you would see in a classical museum. The difference is, it was generated by an algorithm.
We trained GANs on classical portraits and used super-resolution algorithms to produce this high-resolution painting.
We chose the name “Belamy” to make a reference to the creator name of GANs, I. Goodfellow, that roughly translate to “Bel ami” in French. Also, as the signature, we wrote the formula of the loss function of the original GAN model:
We signed with the formula in order to have a clear visual way to show that the portrait was generated by an algorithm.
The portraits in this collection depicts various epochs, and seems inspired from a variety of styles and techniques (oil painting, watercolor…). We carefully picked the artworks to illustrate the creative potential of the algorithm, once trained.
Taking a step back: our perspective
We want to share our perspective on AI for Art.
A new movement is forming: GANism
This new creation process opens up new doors to current world of art, with a new movement François Chollet calls GANism. It offers new possibilities, such as the ability to place an artwork in a given space according to its visual characteristics, and the ability to travel in this space, through an infinite number of altered versions of an artwork.
Accessibility of AI tools
We want to send out an update of the state of the research in AI. Great tools are being discovered by a community of researchers in Machine Learning. Yet, few know what these tools are capable of. By introducing our artworks in a very accessible way (portraits framed that look like something you can find in a museum), we hope to give a view of what is possible with these algorithms.
Further questions
We believe some of the work in AI and Art raises philosophical as well as societal questions. Is an algorithm capable of creativity ? This question has risen many times since computers have been created, but this time, with algorithms like GANs, it takes it a step further.
Many questions are also raised in the field of Art. In contemporary art, the artist has always been at the center of the work, and the tool as a way for him to express, and pass on emotions. Here, the tool is closer to the center of the work, even though the artist behind the algorithm remains the “real” artist. The intention and inspiration comes from the human who designed and used the algorithm. Hence the collaboration between human and machine has never been so close.
This approach is likely to result in the appearance of a new type of art. It will definitely not replace artists, it brings up a new perspective.
Rather than the artwork itself, we believe that the value of this project lays in the debate it can create, and in the exposure of this new tools to the public.
Next steps
We are continuing to experiment with Machine Learning tools for Art. Also, by staying updated with the state-of-the-art of Machine Learning research, we will be hopefully able to spot which new tools can be used for Art.
Conclusion: we want to hear about your perspective.
There is a lot to discuss about. We want to hear your take on the subject, it is a part of our common journey. Don’t forget to keep an open mind !
Contact us at hello.obvious@gmail.com, comment below, tweet us (@obv_ious) or come meet us in Paris.
Hugo Caselles-Dupré.