A mesh of Deren
The training dataset of a neural net must meet the conditions of being quantitatively numerous and at the same time qualitatively heterogeneous and unitary. If the images are very diverse, the GAN will not be able to learn to generalize and generate similar images; if they are too similar, it will only be able to create a number of variations that we would appreciate as limited. Following and subverting this logic we have created a dataset from all the frames in a movie. The images of a film constitute a closed corpus -both in terms of style, as well as historical, as conceptual- and at the same time heterogeneous -if we compare the shots between them- and homogeneous -if we think in terms of the frames composing the shots, which diverge by very subtle and successive variations.
This network has been trained with all the images of Meshes of the Afternoon (1943) by Maya Deren. The result is a network that produces images mainly of abstract tendency but where one can sense a kind of graphic feeling of the original film and the presence of certain obsessive motifs of it (the walk, the figure with a mirror on the face, close-ups, etc.). The routes through the latent space of this network may seem dreams caused by the film; as a kind of representation of the unconscious of the images themselves.