ECCV : Autonomus abstract creation

Abstract art created by a series of AI and python scripts in series, each with one specific task : the first learns from a corpus of nonabstract photographs formed by the author and creates abstract pieces from it. Two AI then make the images bigger and adds learned "texture" to them to give them a more tactile quality. Then an algorythm chooses the "best" pieces based on originality. This last selection was made by the author. 

Geoff Hinton, one of the godfathers of AI, said "The brain's a big neural network. And so, it has to be that stuff like this can work, because it works in our brains. There's just never any doubt about that.", and while it is a simplification, it seemed interesting to try and translate directly the way the human brain works with AI. In art, nothing is created ex-nihilo, but instead the artist "filters" their experiences and knowledge to create artworks, yet they are often visually unrelated to their inspiration. 

Many people working on GAN neural networks and their derivations are trying to optimize them for "realistic" results, but a number of researchers and artists find interest in the artifacts that it produces. Robbie Barrat for example makes weird nude images by training on a corpus of classical nude paintings. I decided to take a slightly different direction, and feed images that were somewhat similar in "tone" (crudely speaking, images that I liked, see images on the right) but with no similarities in terms of subject. The custom GAN algorithm then gave me images with the same "tone" but very abstracted, since it had no proper "shapes" to learn. I found the images to have an aesthetic that I liked, as much as my corpus, and yet they all had something new, because I couldn't make out what it was trying to draw. Somehow, it was truly original.

However the images were small in resolution, and had sometimes visible patterns. To make the images more usable and more pleasing to the eyes, I trained two different implementations of SRGAN to blow up the images, with the same corpus I used to create the images, hoping what the SRGAN learned and used would be coherent with what the GAN "liked" about the corpus. It turned out to work quite well, recreating texture that "worked" with the images, and truly gave it a tactile touch. The use of the two SRGAN allowed for a mix of different textures, as well as higher final resolution. 

Examples of texture added by sizing up the image with SRGANs. 

Once trained, AI can create a very large number of works, therefore I needed something to go through and classify the works. I used a combination of two simple algorythms (ssim and MSE) to extract the most "original" works, and it worked quite well. 
Philosophically speaking, an AI that can create a very high volume of interesting art is a challenge: historically speaking the value of art has always been linked to its scarcity, and maybe we will have to rethink how we value art. The same can be said about the human input: here, while the code was written by humans, I consider the AI to truly be a "partner" in the creative process, in that it creates something that I wouldn't have on my own. 
In terms of research, I can also only adknowldedge all the work of researchers and artists before me. As such, and because the volume of artworks is so big, I intend to create a website where bands, and maybe authors, could "claim" artworks generated by my "AI creator" for either the cover of their albums or books, as a way of giving back to the creative community. Without naming them, I have been following the works of several jurors and their passion gave me the impulse to start my own research, and I hope to be a link as well between previous artists and the ones to come.

Thank you for reading