Brutalist Architecture by a DCGAN

What is brutalist architecture? What are the key features that make a building brutalist? How does brutalist architecture look from the perspective of a neural network? Ultimately, the answers I found were focused more on what features a “brutalist architecture” picture has.

Brutalism has always been one of my favorite styles of architecture. I thought it would be interesting to see what would happen if a neural network were trained to create new brutalist buildings. This kind of neural network used to create new unique images is called a Deep Convolutional Generative Adversarial Network or DCGAN for short.  Using carpeDM20’s DCGAN Tensorflow algorithm, I trained a neural network on a few thousand images tagged with the “brutalistarchitecture” hashtag and the first 500 images that Google returned when I searched “brutalist architecture.” I thought this might be a good place to start because:

  • It is a fairly specific and well-defined style unlike “modernism”
  • It is not too specific like “Brooklyn Romanesque” that there aren’t many pictures
  • There are a lot of enthusiasts that share relevant photos

Nonetheless, a lot of “noise” found its way into my dataset. Some posters related to brutalism were present as well as some selfies with brutalist structures in the background. I hoped that the number of photos used would help separate the wheat from the chaff.

Some random samples from my dataset. Some of the white borders contaminated the final output.

More random samples from my dataset. Some sunset images found their way in.After some cursory browsing, I was satisfied with the results. The set consisted mostly of images of facades, elevations, interiors, and details all in mostly concrete. The images of irrelevant selfies, motivational phrases and sunsets were mostly outliers.

For this exercise, I used Paperspace as my virtual machine. The process has been somewhat positive, but there are a few difficulties I ran into that are still largely unresolved specifically pertaining to their persistent storage service and their Gradient job runner. Ultimately, I found a workaround. I ran the algorithm and got this output:

The array of generated images from the DCGAN. None of these are photographs.

Clearly, these aren’t fooling anyone, but when you squint, they do begin to look familiar at least. I’m particularly impressed with the variety of images. There are some facade details, elevations, and corner details. There are also some that look like they were taken from a distance. A number have blue skies, while others look as though they were shot in black and white. The white borders from the cropping have also made an appearance here in a few pictures.  Some look as though they follow the rules of single-point perspective. Common features of brutalist architecture can also be identified: clean rational window grids, rusticated facades, and clearly defined geometries. Most importantly they look like they feature concrete buildings.

After a closer examination, I noticed that a number of the images featured the same ugly patch of noise on the bottom right. And a few of them look like they are 80% the same. This may be a result of overfitting or an overly diverse dataset.

These all feature the same ugly patch of noise in the bottom right. Some are almost the same picture with one or 2 details changed.

Regardless, I chose a few that I enlarged in photoshop to highlight below:

Brutalist buildings are usually formed with repeated modular elements forming masses representing specific functional zones, distinctly articulated and grouped together into a unified whole.
Another common theme in Brutalist designs is the exposure of the building’s functions—ranging from their structure and services to their human use—in the exterior of the building.
Surfaces of cast concrete are made to reveal the basic nature of its construction, revealing the texture of the wooden planks used for the in-situ casting forms.

Captions provided by wikipedia.

Leave a Reply

Your email address will not be published. Required fields are marked *