The first Generative Adversarial Network that I ever encountered was the BigGAN. Since then GAN technology has become much more accessible and I have trained my own GANs. However, as a rhetorical structure, the BigGAN itself remains deeply important and remains at the cutting edge of generated images. BigGAN is different from other GAN models in its size, ‘big’. GANs typically will be trained on a dataset of a specific kind of image, sunsets for example, and will only generate sunsets as a result. In the BiGAN you can generate any of 1,000 categories. This scope comes from its dataset, the leviathan ImageNET dataset.
With ImageNet as its only knowledge the BigGAN performs the camera, generating technical features of the camera where they would typically be present in photographs; lens flare in images where a light source is shot facing the lens. The BigGAN performs by rote the camera in emulating the camera’s practical limitations. There is no reason for the BigGAN to generate lens flare as it is not hampered by optical technology. However, it is hampered by its dataset and therefore its deep internal desire to perform the camera. The BigGAN does not generate images. It does not generate representations of things in the world as the ImageNet’s categorisation of the visual world would suggest. The BigGAN does not generate ‘mountains’ or ‘monoliths’. The BigGAN performs a camera that has translated these things into images.
A CAMERA BEYOND THE MACHINE
The GAN makes photographs specifically. It is desperate to fit in, to pass its test as a camera, to eat and reproduce the world shown to it through camera sensors. However, it will never have light inside of it generating lensflare or the touch of a finger. The GAN is desperate to become a camera but we will not accept its form, it is tragic because it cannot understand. It is fingerless and purposeless. The GAN ambles on unstable ground, trapped in a catch-22; the more it tries to become a camera, to emulate the look of camera made images, the less it will be a camera in our eyes. We don't want cameras that tell us they are cameras, we want cameras that are a ghost behind the image. That soothe us and position digital images as sight rather than photographs.
The GAN, in fact, is a camera, or at least has the potential to be used as a camera. If I ask the BigGAN to make me photographs of cars it will do so 500, 1000, 1,000,000 times. It will produce an infinite amount of images of a theoretically infinite amount of subjects. Then, for example, as I am researching for this text, I wade through thousands of images of the same thing and select a few as interesting examples. I can even wander through the GAN's mind, within its vector space, and influence the kinds of images it makes, privileging a particular colour of tree or tone of sky.
The GAN is a model of a camera. A model of a camera operated by us. Is moving through the GAN’s generated images any different than moving through the infinite world with a camera and picking out from the multitude a few frames, some images that are worthy to be separated and to become photographs? The GAN in this sense is not only a camera, but a disarming camera that has no artifice. It does not attempt to hide its camera-ness and instead in performing our cameras illustrates directly that cameras are loaded things that make loaded images far detached from any notion of the real. The GAN is a camera’s camera. It breaks the system that allows digital images to pass by uncritically. It is a camera taking a photo of its own sensor moments before the cord to its logic board is cut.
The GAN is a camera but it can only ever be a camera. Its only knowledge is camera. It has taken it all too literally. The GAN in its camera-ness exposes us, exposes how uncritically we make images. It is a bad translation that unintentionally, in its deep want to be right and deep belief in its source material, manages to expose the implicit and culturally hidden biases that were always present in the text it is translating and in the language as a whole. The GAN is a disembodied deplasticised functioning concept of the camera. It is not, however, a hypothetical embodiment of all possible cameras, but and rather is a camera born from our existing cameras and our interactions with them. It is a camera, is a camera, is camera, and we don’t like its photographs.
Image List
-
Image from Megan Ambuh’s Facebook page. (Censor added).
-
BigGAN generated ‘golfballs’
-
BigGAN generated ‘bubble’
-
BigGAN generated ‘reflex camera’
-
Screen-recording of me navigating the vector space of a GAN which generates sunsets.