How This X Does Not Exist Uses GANs to Create Fake Content

This X Does Not Exist is a collection of websites that each generate a specific type of realistic-looking fake content on demand — faces, cats, rental listings, resumes, and more. Most of them rely on generative adversarial networks (GANs), a type of machine learning model that learns to produce new samples that resemble its training data. The best-known example, This Person Does Not Exist, uses a specific GAN variant called StyleGAN to generate photorealistic human faces. Not every site in the collection uses a pure GAN, though: at least one uses a recurrent neural network (RNN) for text, and another uses a simpler model that isn't deep learning at all.

What a GAN actually does

A GAN trains two networks against each other:

• The generator takes random input (often called a noise vector) and produces a synthetic sample — an image, in most of these sites.
• The discriminator looks at samples and tries to tell whether each one is real (from the training data) or fake (from the generator).

During training, the generator is rewarded for fooling the discriminator, and the discriminator is rewarded for catching fakes. Over many rounds, the generator gets better at producing outputs the discriminator can't distinguish from real data. Once training is done, the generator alone can be used to create new samples — which is what you see when you load one of these sites.

The practical result: the site doesn't store a fixed library of images and show you one at random. It runs the trained generator to create a fresh output, which is why refreshing the page usually gives you something new.

The sites in the collection and the models behind them

Two details worth noting from the source description:

• This Person Does Not Exist is described as the site that started the trend, built with StyleGAN and created by Phillip Wang.
• This Resume Does Not Exist is explicitly not a GAN — it uses TextgenRNN, a recurrent neural network specialized in text generation. Its headshots come from This Person Does Not Exist.

Why StyleGAN matters for the faces

StyleGAN's contribution is control over style at different scales. Instead of generating an image in one pass, it separates coarse features (face shape, pose) from fine details (skin texture, hair). That separation is a big part of why the faces look convincingly photographic rather than blurry or distorted. The source description notes that This Person Does Not Exist "had the tech community buzzing" and inspired many of the other sites in the collection.

Limitations you'll actually notice

The source description is candid about failure modes:

• This Cat Does Not Exist: "Once in a while the cats have visual deformities due to imperfections in the model." These artifacts are a normal consequence of imperfect GAN training, not a bug in the website.
• This Emotion Does Not Exist: the description frames it as a demonstration of a simple model's ability to interpret emotions from facial expressions — a different task from generating images.
• This Vessel Does Not Exist: the site draws a parallel between a GAN's generator/discriminator pair and an apprentice/master pair in ceramics, which is a useful mental model for why adversarial training produces skilled output.

How to use this understanding

If you load any of these sites, expect a new synthetic output on each refresh, generated by a trained model rather than pulled from a gallery. If you're evaluating whether a given site uses a GAN, check the description: the collection itself distinguishes GAN-based sites (faces, cats) from RNN-based ones (resumes) and simpler models (emotions). That distinction matters if you're trying to understand what kind of model produced a given output, or if you want to reproduce the approach yourself.