Art Brought To Life With The Help Of Machine Learning

As a graphic artist and computer nerd, I enjoy testing out the juncture between art and science. Lately, I have been experimenting with applications of Generative Adversarial Network (GAN) based art and its ability to to generate images of people that go past the uncanny valley into the realm of the real. GANs are a class of machine learning frameworks in which two neural networks work against each other in a zero-sum game in which one agent’s gain is another agent’s loss: a “generative network” generates candidates that look “real” while a “discriminative network” evaluates the candidates and tries to figure out what is real and what was generated. The results are quite impressive. My work uses an implementation of StyleGAN (see A Style-Based Generator Architecture for Generative Adversarial Networks, https://arxiv.org/abs/1812.04948). I feed images of the statue, painting, or drawing that I want to reimagine as a “real” person into the software (I sometimes use multiple inputs and combine them), tweak the available parameters, and finally do post-processing work in Photoshop. The results below show you what is possible with this incredible technology. The left part of each image shows a famous statue or painting, with the right showing the results of my work. The “real” images give you an idea of what the model behind each work of art may have looked like in real life.