Computer Vision Art Gallery

ICCV 2019, Seoul, Korea

Robbie Barrat

Correction After Artemisia Gentileschi "Lucretia", (2019)


Since the summer of 2018, I have been in confrontation with a traditional French painter; Ronan Barrot. We produced many works together as part of a conversation and argument about artificial intelligence as an artistic medium and tool. While “Correction after Artemisia Gentileschi, “Lucretia”” is not a part of the work we produced together, the process it is heavily influenced by the way I observed Ronan work in his studio. When Ronan was not satisfied with a part of one of his in progress paintings, he would cover up the offending area with bright orange paint, and fill it back in to (in his words) “correct” it. Ronan also would “correct” some of the GAN generations I produced during our confrontation by painting directly over parts of them.

I was deeply interested by this; and decided to try and teach a neural network to do something similar. By repeatedly covering up areas of paintings (specifically nude portraits) and having a neural network reconstruct the obscured areas; the neural network is doing a very technical and obvious task of inpainting (a form of reconstruction/repair) - but also adjusting the painting incrementally to better fit the neural network’s internal representation of what a nude portrait looks like. This act of correcting the painting to be in line with the neural network’s internal representation of nude portraiture only emerges throughout lengthy repetition of inpainting/reconstruction; and oftentimes the entire process results in a corrected image that hardly resembles the initial one.