PAPERS-04-002 · Paper — Generative Adversarial Networks (GAN)

The paper

The 2014 paper "Generative Adversarial Networks" by Ian Goodfellow and colleagues, presented at the NIPS conference, is a landmark because it introduced a completely new way to train generative models without needing explicit likelihoods, opening a path to high‑quality image synthesis.

Before this paper

Before GANs, generative models depended on explicit probability formulas that were difficult to train and to sample from. Their training often got stuck, and the images they produced tended to be blurry. Researchers struggled to create sharp, realistic samples with likelihood‑based approaches. This limitation formed a major wall for progress in generative modeling.

The key idea

The core insight was to pit two neural networks against each other: a generator that tries to create fake images and a discriminator that tries to tell real images from fakes. By training them together in a minimax game, each network continuously improves. This adversarial setup removes the need for an explicit likelihood function.

How it works

First, the generator receives random noise and transforms it into an image. The discriminator then evaluates that image, assigning a score that reflects how real it thinks the image looks. The discriminator also evaluates real training images and learns to give them high real scores. The generator’s loss is based on how well its fakes fool the discriminator, while the discriminator’s loss reflects its ability to distinguish real from fake. Gradients from these losses update both networks, and the process repeats, with the generator getting better at producing realistic images and the discriminator getting sharper at spotting fakes.

See it — the forger vs the critic

flowchart LR
  N["Random noise"] --> G["Generator<br/>makes a fake"]
  G --> D["Discriminator<br/>real or fake?"]
  R["Real images"] --> D
  D --> G

What they showed

The authors showed that GANs could produce sharp, realistic samples that earlier likelihood‑based models could not, demonstrating a clear qualitative jump in image quality and sparking a wave of ever‑more‑photorealistic synthesis.

What it changed

GANs launched the modern era of generative modeling and introduced the powerful concept of adversarial training. This shift inspired countless follow‑up works that pushed image synthesis to new heights. The approach reshaped research priorities, moving focus from likelihood estimation to adversarial objectives, and it quickly spread to many applications beyond images. Next: Variational Autoencoders (VAE).

The one-line takeaway

Adversarial training lets two networks learn together to create surprisingly realistic data without explicit probability models.

Read it / meet it

Reading the original GAN paper gives a clear view of the simple yet profound idea that started a revolution. Today, every "this face does not exist" demo and many modern image tools trace their roots back to this work, so exploring it connects you directly to the technology you see every day.