Exploring the notion of the Baudrillardian referent in the “real object” of Graham Harman’s object-oriented ontology, Ciel is a project in which a Deep Convolutional Generative Adversarial Network (DCGAN) is trained to produce a selection of sky-like images.

The process itself involves simultaneously training a generator and a discriminator, coded in Python, the former trying to get better at generating images that look real, and the latter trying to get better at distinguishing real images from the fakes.

Ultimately, the two conflicting systems achieve an equilibrium when the discriminator can no longer distinguish the real images from the fakes. The video shows a transition through the training epochs of the generator, with samples taken at various stages during the training process.