The Latent Image

If you believe, as I do, that human artists have covered most of the possibility space, the question becomes, what can you still do as an artist?¹ (Mario Klingemann)

If recent years have witnessed the mass adoption of generative AI as a creative tool, Fellowship’s forthcoming show of early works by Mario Klingemann captures the artist’s investigations of generative modelling before latent space became public space. Projects such as SketchMaker (2007-2014), Neural Abstractions (2016-17), CycleGAN Makeover (2017), and Chicken or Beef (2018) offer a window into the artist’s evolving systems before public models such as StyleGAN and BigGAN, which used GANs (generative adversarial networks), began overtaking his research in their capacity to generate plausible likenesses of human figures. 

By spotlighting this seminal decade of the artist’s career, the exhibition proves that the hand coding of the Flash community precipitated a new ontology of posthuman collaboration.

Where 2018 represented a “Cambrian explosion” for Klingemann, it was also the moment a marginal province of academic research became an expanded art of machine learning. For while Robbie Barrat’s AI Generated Nude Portraits (2018) — infamously disregarded at Christie’s Art+Tech Summit — were christening a new age of digital ownership, Klingemann’s use of custom upscaling known as “transhancement” ensured that his own brand of disfiguration was hi-res enough to hang on a wall. In this respect, he followed Harold Cohen (1928-2016) in making machine-made art fit for public consumption. 

Klingemann also posed questions of neural networks that Cohen had once asked of classical symbolic AI when developing his art-making program AARON. But whereas AARON was an expert system that encoded the heuristics of drawing, composition, and formal decision-making into a procedural framework, Klingemann was more interested in navigating the statistical terrain of latent space: a world invisible to human eyes. The fact that his investigations ultimately begot Botto — a “decentralized autonomous artist” whose practice evolves via community curation — underscores his unconscious affinity with Cohen as well as the relational potential of neural networks. 

The works in this show anticipate that move toward relational creativity while remaining distinct from Botto in one crucial aspect: rather than externalizing the process of curation to a decentralized community, Klingemann is curating latent space himself. 

Instead of negotiating the machine’s autonomy in public, these works are private explorations of the border territory between human and nonhuman vision. 

In the ideal world I would never have to produce any output. I would just enjoy the process… For me, the outputs are just a receipt that the system is working. (Mario Klingemann)

In admitting his preference for process over end product, Klingemann aligns himself with Sol LeWitt, for whom the execution of the artwork was a “perfunctory affair.”² LeWitt’s Wall Drawings, which he produced from 1969 until he died in 2007, relied on a set of rules to guide technicians in executing the final outcome. Following LeWitt’s logic, the production of the artwork became secondary to its conceptual framework. 

In today’s era of prompt engineering, when creativity is the new productivity, Klingemann’s privileging of the system over the output feels prescient. It is also a consequence of his background in graphic design, where the act of execution was distributed across different technical processes. His willingness to use “stolen photography” as the basis for his flyer designs for the Munich techno scene of the 1990s reinforces the importance of appropriation to his practice, which ultimately fueled his interest in neural networks as machines for repurposing media. 

Yet Klingemann was already coding in the 1980s, and his early experiences of programming in BASIC with a Commodore C64, Casio PB 100, and a SHARP PC 1500 grounded his art in procedural logic. Indeed, he regards his Photoshop plugins from the ’90s as his first works of generative art. Falling in with the Flash community, which became a nexus for creative coding in the early 2000s, Klingemann befriended Jared Tarbell, whose algorithmic explorations dynamized the field of computational aesthetics for the internet age. According to Klingemann, SketchMaker (2007-2014) was a direct outgrowth of the Flash ethos, serving as a search mechanism for the discovery of emergent aesthetic possibilities. 

Consisting of modular algorithmic components, SketchMaker echoes early rule-based generative art in seeking to “tame randomness” in search of aesthetic “interestingness.”

It was in his Neural Abstractions that Klingemann made the leap from procedural generative art into the latent space of machine learning. Experimenting with Plug-and-Play Generative Networks (PPGN), he fed the raw, low-resolution outputs into a second pix2pix-based model, producing “proto-shapes” that emphasised errors of representation. The indeterminacy of the resulting images, which veer between figuration and pure-color abstraction, satisfied the demand of modern easel painting while expressing the uncertainty of a machine learning to see. They also unsettled human spectators not used to “seeing” latent space. Of course, the white cube of the gallery has always been in some sense a black box, whose aesthetics encode the politics of the marketplace. But at a time when corporate image generators are still laden with bias, Klingemann’s Neural Abstractions surrender human (and implicitly Western) domination of optical experience.

With Web2 announcing a new age of algorithmic identity, where social media platforms acculturated users to saleable norms, human imagination was up for grabs. Klingemann’s riposte was to cross-fertilize male and female types using a new model called CycleGAN. Trained on open archives of rather solemn 19th-century portraits, the resulting hybrids are a beguiling mix of sculptural presence and painterly elusiveness. Yet their romantic disfiguration also parodies the pernicious logic of “facetuning” by which online identities are commodified according to capitalist imperatives. 

Analysis of Klingemann’s early GAN-generated works often stresses their visual similarity to Francis Bacon’s uncanny bodies, whose distended features seem to externalize internal torment. But according to Klingemann, this is really a projection of the viewer. The real reason works such as Chicken or Beef ended up in Bacon’s “neighbourhood” was because the model was trained on pornographic datasets while still groping toward realism. 

Like all his works with machine learning, they carve out space between human and machine comprehension, but by emphasizing errors in representation they replace canonical genres with a state of aporia.

Klingemann’s works to 2018 trace a tectonic shift from hand coding to system building with machine intelligence. Shaped by generative error and statistical bias, his latent images resist resolution by human viewers while evidencing the ways human experience defies machine understanding. Trained on fragments of the Western imagination, these series highlight the “invisualities” inherent in big data along with the violent potential of nostalgia.³ As one of a small number of artists who trained their own models in the decade before generative AI went mainstream, Klingemann’s works remain vital documents of the moment machine learning unsettled the human imagination.