Denmark's First GAN Artworks - AI Art History

When did the first GANs come out?

2014

Goodfellow et al. publish the GAN paper

Ian Goodfellow and his colleagues at the University of Montreal introduce Generative Adversarial Networks at NeurIPS. The idea: two neural networks competing against each other - one generates, one discriminates. The results are blurry and low-resolution, but the concept changes everything.
2015-16

DCGANs make it practical

Deep Convolutional GANs (DCGAN) prove that the architecture can generate recognisable images. Researchers start sharing code on GitHub. It is still dense, poorly documented, and mostly runs on university compute clusters.
2017

pix2pix, CycleGAN, and ProGAN arrive

Image-to-image translation becomes possible. NVIDIA's Progressive GAN (ProGAN) generates the first convincing high-resolution human faces. A small global community of artists begins experimenting. There is no tutorial scene, no community Discord - just arXiv papers and raw code repositories.
2018-19

First GAN artworks made in Denmark - built from scratch

Working independently, I build a GAN from scratch in TensorFlow and Python - no pre-trained weights, no off-the-shelf tooling. The results are among the first GAN artworks created in Denmark. StyleGAN hasn't been released yet. The Christie's AI art auction shocks the art world this same period. Almost nobody in Denmark knows what a GAN is.
2019-21

StyleGAN and the mainstream turn

NVIDIA's StyleGAN produces photorealistic faces with unprecedented control. "This Person Does Not Exist" goes viral. AI art starts appearing in galleries. The era of accessible GAN tooling begins - but most of the exploratory, raw work was already done.

What it was actually like

The first time I trained a GAN, the outputs were noise. Literally - just static. That was expected. You were supposed to babysit the training process for hours, sometimes days, watching the generator and discriminator find their equilibrium. If the discriminator got too good too fast, the generator would give up and produce the same image over and over, a failure mode called mode collapse. If neither won, you got noise forever. Getting anything coherent out felt like coaxing something alive.

Most people had never heard the word "AI" in the way we use it today. Not in Denmark, and honestly not anywhere outside machine learning research circles. If you told someone you were making art with a neural network, they would ask what that was. Not in a dismissive way - they genuinely didn't know the concept existed. There was no cultural reference point. AlphaGo had beaten a Go world champion two years earlier and that had briefly made news, but it felt abstract and distant. Nothing had made neural networks feel relevant to everyday life.

The research community was different then, too. If you searched for "neural networks" or "AI" on YouTube, you did not get what you'd find today. No clickbait thumbnails, no "AI will replace you" anxiety loops, no five-minute explainer videos optimised for outrage. You got lectures. Actual university lectures - Geoffrey Hinton teaching at Toronto, Andrej Karpathy's Stanford CS231n course, Yann LeCun talking at ICLR. Dry, long, genuinely informative. The audience was people who wanted to understand the thing, not people who wanted to have an opinion about it.

There was something clarifying about that. The field hadn't been colonised by narrative yet. A GAN paper was just a paper. You read it, you tried to implement it, it broke in twenty places, you asked on a GitHub issue thread and waited three days for an answer. The gap between an idea and a working result was enormous - and that gap was where the interesting work happened.

Training on consumer hardware meant overnight runs, unexpected crashes, and outputs that felt genuinely strange because nobody had established a visual vocabulary for what GAN art was supposed to look like. There were no aesthetic norms to conform to or rebel against. The images that came out were sometimes deeply uncanny, sometimes unexpectedly beautiful, and always a negotiation between the dataset you'd curated and a process you only partially understood.

I don't think that uncertainty made the work worse. I think it made it more honest.

Frequently Asked Questions

When did the first GANs come out?

Generative Adversarial Networks were introduced in June 2014 by Ian Goodfellow and colleagues at the University of Montreal in the paper "Generative Adversarial Nets," presented at NeurIPS. The first GANs produced low-resolution, blurry outputs - practical high-quality image generation came with DCGAN (2015), ProGAN (2017), and StyleGAN (2018-2019).

What is GAN art?

GAN art is imagery generated by a Generative Adversarial Network - a neural network architecture in which a generator learns to produce realistic images by competing against a discriminator that tries to tell real from fake. The artist's role is in curating the training data, shaping the architecture, and selecting outputs. It is a collaboration between human intent and emergent machine behaviour.

Who were the first AI artists?

The first artists working with GANs emerged around 2016-2018. Internationally, names like Robbie Barrat, Helena Sarin, and the Obvious collective (whose work sold at Christie's in 2018) are among the early figures. In Denmark, this work was being done in parallel, independently - outside any institutional framework and largely unknown to the wider public.

What hardware did early GAN training require?

Most early GAN work outside academic settings ran on consumer GPUs - NVIDIA GTX 1080 or similar. Training a single model could take 12-72 hours. Cloud GPU access existed but was less standardised and more expensive than it is today. Efficient mixed-precision training wasn't widely adopted yet, so VRAM was a constant constraint.

Is GAN art still relevant now that diffusion models exist?

Diffusion models (Stable Diffusion, DALL-E, Midjourney) have largely replaced GANs as the dominant image generation architecture since 2022. But GAN work from 2017-2021 has a distinct visual character - artefacts, latent space interpolations, the specific uncanniness of ProGAN faces - that diffusion models don't replicate. Early GAN art is a distinct historical category.

The first GAN artworks in Denmark

When did the first GANs come out?

Goodfellow et al. publish the GAN paper

DCGANs make it practical

pix2pix, CycleGAN, and ProGAN arrive

First GAN artworks made in Denmark - built from scratch

StyleGAN and the mainstream turn

The Artworks

AI-Generated Portraits - Stylized & Abstract

GAN Series - Neural Network Output

What it was actually like

Frequently Asked Questions

When did the first GANs come out?

What is GAN art?

Who were the first AI artists?

What hardware did early GAN training require?

Is GAN art still relevant now that diffusion models exist?

Working with AI since before it was a trend