**Neural Abstractions (2016–17)**
Neural Abstractions represents Klingemann’s breakthrough into true AI-assisted image generation, using early deep neural networks to summon surreal new visions. Created in 2016–17, these works grew out of experiments with Plug-and-Play Generative Networks (PPGN), an innovative technique that predates GANs, allowing an AI to generate images by maximizing the outputs of a classifier rather than by direct synthesis. Instead of training a network to draw per se, Klingemann ingeniously trained a suite of custom classifier models (on datasets he curated himself) and then set a generative process to “please” those models. In practice, he fed the network a target concept, not via text prompt, but via a classifier node such as “portrait” or “map”, and the system iteratively adjusted a noise image until the classifier was highly activated.
Rejecting the generic categories bundled with the original PPGN code (which skewed towards dogs and everyday objects), Klingemann plugged in his own bespoke categories. He hand-sorted tens of thousands of images, from vintage album covers to 19th-century book illustrations to even erotic photographs, into conceptual groupings, training neural nets to recognize these themes. In doing so, he effectively hacked the generative process to follow the peculiar contours of his personal datasets. The resulting images, which he termed “Neural Abstractions,” are richly ambiguous. Hints of familiar shapes, a silhouette that could be a reclining nude or the glint of an eye, dissolve into painterly washes of color and texture. The AI seems to grope toward forms without ever fully resolving them, offering a “base understanding of composition and interesting colors” while birthing proto-figures that hover on the brink of recognition.
Because the raw outputs were modest in size and clarity, a limitation of 2016-era networks, Klingemann devised a novel post-processing step that would become a hallmark of his style: transhancement. This custom technique fed the low-resolution neural outputs into a secondary pix2pix-based model that upscaled and hallucinated extra detail, layer by layer. Unlike traditional upscaling which merely smooths or interpolates, transhancement exaggerates the AI’s own artifacts, creating “beautiful convolutional artifacts” and razor-sharp edges that amplify the image’s uncanny character. By repeatedly doubling an image’s size, Klingemann transformed a 256px blurry apparition into a large, crisply freakish tableau.
The artifacts - glitchy striations, bizarre textures - are not flaws but the very soul of the work, imbuing it with an otherworldly, hallucinatory quality. In Neural Abstractions, one discerns a machine dreaming aloud, as if it were trying to reconstruct reality from only half-remembered shapes. This series solidified Klingemann’s status as a pioneer of AI art, establishing an aesthetic of ambiguity and unpredictability that remains “uniquely computational yet deeply human.”

Neural Abstraction #33

From

Neural Abstraction 2017

Ξ1.111

Description

Neural Abstractions (2016–17) Neural Abstractions represents Klingemann’s breakthrough into true AI-assisted image generation, using early deep neural networks to summon surreal new visions. Created in 2016–17, these works grew out of experiments with Plug-and-Play Generative Networks (PPGN), an innovative technique that predates GANs, allowing an AI to generate images by maximizing the outputs of a classifier rather than by direct synthesis. Instead of training a network to draw per se, Klingemann ingeniously trained a suite of custom classifier models (on datasets he curated himself) and then set a generative process to “please” those models. In practice, he fed the network a target concept, not via text prompt, but via a classifier node such as “portrait” or “map”, and the system iteratively adjusted a noise image until the classifier was highly activated. Rejecting the generic categories bundled with the original PPGN code (which skewed towards dogs and everyday objects), Klingemann plugged in his own bespoke categories. He hand-sorted tens of thousands of images, from vintage album covers to 19th-century book illustrations to even erotic photographs, into conceptual groupings, training neural nets to recognize these themes. In doing so, he effectively hacked the generative process to follow the peculiar contours of his personal datasets. The resulting images, which he termed “Neural Abstractions,” are richly ambiguous. Hints of familiar shapes, a silhouette that could be a reclining nude or the glint of an eye, dissolve into painterly washes of color and texture. The AI seems to grope toward forms without ever fully resolving them, offering a “base understanding of composition and interesting colors” while birthing proto-figures that hover on the brink of recognition. Because the raw outputs were modest in size and clarity, a limitation of 2016-era networks, Klingemann devised a novel post-processing step that would become a hallmark of his style: transhancement. This custom technique fed the low-resolution neural outputs into a secondary pix2pix-based model that upscaled and hallucinated extra detail, layer by layer. Unlike traditional upscaling which merely smooths or interpolates, transhancement exaggerates the AI’s own artifacts, creating “beautiful convolutional artifacts” and razor-sharp edges that amplify the image’s uncanny character. By repeatedly doubling an image’s size, Klingemann transformed a 256px blurry apparition into a large, crisply freakish tableau. The artifacts - glitchy striations, bizarre textures - are not flaws but the very soul of the work, imbuing it with an otherworldly, hallucinatory quality. In Neural Abstractions, one discerns a machine dreaming aloud, as if it were trying to reconstruct reality from only half-remembered shapes. This series solidified Klingemann’s status as a pioneer of AI art, establishing an aesthetic of ambiguity and unpredictability that remains “uniquely computational yet deeply human.”

Mario Klingemann

Neural Abstraction #33

From

Neural Abstraction 2017

Ξ1.111

Description