Science

“Moebia”, from text to drawing with an algorithm – Sciences et Avenir

This is a graphic novel released this summer with a visual style reminiscent of cartoonist Moebius. In fact, the work is called Mebia. This is a story of waiting, in which a small group of living forms arrive on the planet Earth, from which people have disappeared. “Only giant crows, scorpions and mysterious nuclear tanks remained…” the authors say. The authors are, to be sure, a bit specific. Moebia is a project of a French creative studio founded in 2018, u2p050. His works, films, exhibitions revolve around the theme of “the growing digitization of the world and the impact of digital technology on a philosophical level,” the co-founders point out. But in fact, 70 pages of the book were adapted into drawings from text using the VQ-Gan artificial intelligence algorithm.

Two competing algorithms

In particular, it is the antagonistic generative network (GAN), a type of tool invented in 2014, widely used in artistic matters and repeated several times in Science et Avenir. The method involves running two algorithms against each other to create content without human intervention. If it is not the original database created to train one of the two algorithms, which, according to this, will “evaluate” the production of the second.

VQ-Gan, developed in 2020 by three researchers at the University of Heidelberg in Germany, is a text-to-pixel conversion algorithm. In other words, it transforms the text given to it into an image. Specifically, the u2p050 team used a Google extension developed by AI generative art researcher Katherine Crowson. It is possible to use different versions of the algorithm, depending on whether it is trained with a specific image library (Wikiart, Openimages, ImageNet, Coco, Faceshq, Sflckr), resulting in different results.

Scorpions missing from original text

In this case, for Moebia, U2p050 studio decided to use SFlckr as a basis. “First, the whole text was written, then we put the sentences one by one in the “text in pix”, explains Michel Nerval, one of the authors in collaboration with Jeanne Knoplioch, Pauline Broquet, Zoe and Philippe Ngo, and the adaptation was carried out gradually, with sometimes small additions. Scorpions, for example, did not appear in the original story. That was one of the goals of the project: to see what the AI ​​has to offer visually, and perhaps follow its choices.

Mobia was printed on paper in an edition of 200 copies. (Credits: u2p050)

However, human intervention is not zero. With each piece of text entered into the machine, the algorithm generates an infinite number of images on the fly. It is necessary to check how many visuals the result can become exploitable, and then adjust this threshold so that the generation stops automatically. Out of 2000 images, 60 had to be kept. The algorithm was also modified to increase the number of pixels and retouch colors. “This is a very exciting process based on feedback between machine and human,” continues Michel Nerval. “If we didn’t like the image, we added or removed a word from the text.”

The book, originally available digitally, was printed on paper and published by the studio in an edition of 200 copies.

Back to top button

Adblock Detected

Please consider supporting us by disabling your ad blocker.