Data Moiré

The larger the dataset, the better the generated images in terms of photographic quality. The diversity of content within the dataset also was found to impact the generative capacity of GANs directly. The smaller variation in the dataset produced a more focused and better image, while the more variation in the dataset created more diverse results, but usually with less accuracy.

The subject of dataset was an important reflection of the general definition of the concept of dataset, a general definition that must have been as close as possible to swissness. Unfortunately, Swissness and also a vague set resulting in a part of the notion of several countries and landscapes. A necessity to find a correct limit.

The result of his interpolations are entities no longer based on a form in a general way but rather on vectors of the directions and forces of the networks that transcend the essence of what is the Swissness.

In addition, other than the classical use of DCGAN, part of the work was also the exploration of the possible abstraction by DCGAN notably by merging equivalent architecture {512, 512, 3} and {512, 512, 3} but also with different architectures {512, 512, 3} and {256, 256, 3}. And the GAN export has equal tridimensional size in order to generate objects and geometrical shapes in a matrix {32, 32, 32}.

This interpolation asks about the Swiss ornament, the position of the openings is principally what the building wants to tell us from the inside and from the outside to the inside. The openings as soul and personalization of the building as ornament. Sometimes aligned, staggered, grouped, it forms an array network that gives personality to a concrete cube. At the heart of this interpolation it is not only the window that is represented but the influence of the window on the abstract form of the facade and therefore the way in which it decorates it.

This interpolation is chosen for its aspect of pattern in which we can still read a vertical array network inside, we could see a moiré so hidden and homogeneous that creates a final pattern. The question of this facade having a personality with the ornament given by the pattern also influenced the choice of this interpolation but also the unfinished aspect that it proposes. The project can continue in an infinite way, it has no floor, no roof, no limit and does not prevent it from propagating, transforming this interpolation in an instant and fragment of a larger entity.

GAN Image
GAN Analysis

The axes, the balcony, the shutters are all influenced by the opening and make an important part of the ornament of the Swissness facade. This building has also no limits, it can be extended infinitely.

The first choice made was not to represent the GAN as an object but an abstract artifact.  There are made by simple transformations: in a first step a fusion of the 3 layers of the RGB — XY to XY XZ YZ — and in analyzing the contrasts under a binary logic, in order to obtain a boolean voxel matrix with three objects. From each boolean value an object is created: 0 null, 1, full, 2 existing, 3 empty cube of X dimension canceling 0 and 1. This creates the paths through the positive voxel, creating the shape.

Translation result

Keywords: GAN, Artificial Design, Swissness, Housing

Master project student: Enzo Arioli Bildstein

Studio tutors: Christina Doumpioti, Christoph Holz, Mikhael Johanes, Frederick Kim, Gianna Ledermann
Professor: Jeffrey Huang

Image credits: Enzo Arioli Bildstein, MxD