Common Daemoniator

Description (in English)

It has been said that deep learning is a modern form of alchemy: researchers are using deep neural networks to transmute digital matter and are successfully generating new meaning from random noise without always being able to fully explain how or why a particular process works. By assembling and training models they create high-dimensional latent spaces whose inner workings and rules are often too complex to be fully comprehended so their exploration becomes a mixture of try-and-error, gut feelings and personal belief systems.

StyleGAN2 is one of the latest generative adversarial networks that is able to create realistic human faces of persons that do not exist. It is a 21st century alembic in which homunculi are synthesized, capable of showing the face of every human being that has ever been born or might be born in the future, from every age, gender or race. Every face is made from a series of 512 numbers which control every aspect of it: the direction it looks to, the motion of its eyes and mouth as well as its emotions that we perceive through the complex interplay of its facial features. Every one of these numbers is like the string on a marionette that controls various aspects of the virtual being but the complex nature of the model's latent space makes it impossible for us to really understand the rules behind it. There are no instructions and no map for how to navigate through latent space - we have feel our way through it step by step using makeshift tools that we adapt with every bit on knowledge we gain in our explorations.

Common Daemoniator is a latent laboratory shared by all visitors in real time and allows everyone to conduct alchemistic experiments in a virtual petri dish. The outcome is in constant flux and the result of the common actions of the group. Depending on how the strings are being pulled we might encounter Aphrodite or a daemonic creature.

Source: https://esc.mur.at/en/werk/common-daemoniator

Situation machine vision is used in

Authored by