Latentscapes #3, 2018

“In 2017, without realizing its significance then, I started training neural networks from scratch solely on my own artwork - the thing that would come to define my artistic approach. A handful of popular image synthesis models available at that time didn’t work well with small datasets so I started experimenting with obscure model architectures which resulted in a few interesting series like this one where I trained the network on my travel photography - coining the word Latentscape. The problem was that the generated images were tiny. Not giving up I coded a computational pipeline of chained models, gradually upscaling the images - calling this a “poor woman super-resolution”. Coincidentally Jason Bailey wrote an article in 2018 “Helena Sarin: Why Bigger Isn’t Always Better With GANs And AI Art" where one can find more details on my artistic philosophy and techniques working with small datasets using your own small compute.” Artwork spotlight by Luba Elliott: Helena Sarin’s work has always stood out amongst artists working with generative adversarial networks (GANs) in the late 2010s because of her focus on craft. Meticulously stitching together diverse mediums and GAN algorithms, Sarin’s work espouses a unique aesthetic, treating AI artwork-making as a craft-based process that highlights the artist’s mark rather than hiding it. This places her closer to the lineage of printmakers and German Expressionists such as Erich Heckel and Conrad Felixmüller, whose gestural marks define not only the form of the subjects, but also the mood of the piece with their immediate intensity. However, the tools used by Sarin are digital. With a background in software engineering and a parallel creative practice in the applied arts, Sarin’s art was predominantly analog until she came across GANs, an image generation technology that consists of two neural networks, a generator that produces images in the style of a particular dataset and a discriminator, which determines whether the generated images are real or fake. The interplay between these two systems typically results in high-quality images and offers increased dialogue between artist and machine, with the need to constantly tweak the weights, parameters and saving intervals of the model in order to achieve the desired results. The unpredictability of these systems and the rapid pace of their development in a race to beat state of the art consistently offered new models to experiment with, each with its own affordances, especially when applied to small datasets. In an era where bigger is better, Sarin made a name for herself by honing the craft of training smaller, less-computationally intensive models on her own datasets in order to exercise a greater creative control over the aesthetic of the output. For Latentscapes, Sarin’s dataset of choice is a collection of her travel photography across the US - seascapes and mountains from the East Coast to the South West. This was then used as training data for SNGAN-with-projection, an obscure GAN model capable of training and generating multiple classes of images. The resulting selection became Latentscapes, a term coined by the artist as a portmanteau of ‘latent space’ and landscape. Latent space is a mathematical representation of compressed data, in which similar data points are grouped together. Sarin becomes a traveller across this new terrain, her programming skills replacing the camera as she captures specific samples that catch her eye in-between categories. These snapshots document the latentscape as a machine environment not bound by explicit rules of object representation and arrangement. Veering towards abstraction, these images separate water, mountains, clouds and trees through textured brushstrokes and marks of varying levels of detail and direction, creating energetic scenes that are both reminiscent of natural landscapes and their artistic representation through oil painting, with visible application of palette knife on wet paint and giant brushstrokes. Yet at the same time these snapshots communicate an alien perspective, where the sky is both above and below the approaching wave of the forest, such unexpected compositions reminding us that we are travelling within a machine’s understanding of the environment.