The curator's reviewRefresh the page, meet a stranger who has never existed. Every face is dreamed whole by a GAN: pores, stray hairs, a life story in the eyes, and none of it real.
A single face fills the page. Refresh: another face. Refresh: another. None of these people exist. Philip Wang put StyleGAN - NVIDIA's then-new face-generation network - behind a bare domain in February 2019, and in doing so taught more of the public what a GAN was than any paper or press release ever managed.
The faces reward attention. Skin texture, stray hairs, catchlights in the eyes - flawless. Then you drift to the edges: earrings that do not match, backgrounds that melt into paisley static, companion faces at the margin dissolving into nightmare. The network learned the face with total dedication and learned everything around the face like a rumor. Those failure modes became the founding literacy of deepfake detection: check the earrings, check the background, count the teeth.
The site is also a historical marker. 2019 was the moment synthetic media crossed from research curiosity to public fact, and this page - one URL, no interface, infinite strangers - was the crossing. Every this-X-does-not-exist site that followed (cats, words, cities, feet) is its descendant.
You would nod at these people on the street. That is the part that stays with you: not that the machine fails at the margins, but that it succeeds so completely at the center. A haunted mirror, kept politely refreshing since 2019.
