A Walk Through Latent Space: Making Art With Neural Networks

by Gregory Wieber

These are not a paintings. At least not in the normal sense of the word. They are machine hallucinations, produced by a neural network that was trained on a dataset of Fauvist art — a group of French artists at the turn of the 20th century, who influenced modern masters like Picasso.

The Fauvist school was itself hallucinatory, with its bright colors and abstract compositions, fueled by absinthe and the natural beauty of Provence. This style lends well to a neural network generative approach, which produces images that feel like a hallucination, or a dream. What are these places, these people, these paintings? They feel familiar, like the neural net has captured not just an aesthetic, but memories of the past.

The number of Fauvist paintings is limited, so transfer learning was used — wherein a model that had already been trained on a wide range of generic art (stylegan-paintings) was used as a jumping-off point for continued training with the Fauvist image dataset. In a sense, this process mirrored the Fauves themselves — building on a base of classic art training, and riffing off the styles of their contemporaries.

Like this work? Follow me on instagram for more.