Purpose: This website is used to generate random, realistic human faces with one click, and these faces belong to people who do not exist. The generated results can be downloaded and are suitable for personal use.
Main capabilities: The website is driven by the StyleGAN neural network developed by Nvidia in 2018. The network consists of two competing neural networks, one generating images and the other judging whether the images are real, trained until the generating network continuously deceives the discriminator network. The generated faces include men, women, and children, and the images are 1024x1024 pixels with no watermark. The generated results update every 2-3 seconds.
How to use: After users visit the website, a random face has already been generated. If they do not like the current face, they can refresh the page to get a new face. If they see the same face, they should wait a few seconds before refreshing. Users can directly download the image.
Applicable scenarios: It can be used for personal projects that require avatars of non-real people, such as design, presentations, or testing. For personal use only.
Limitations: The website only provides generated results and does not provide a service for identifying fake faces. Identifying fake portraits is almost impossible; 90% of fake faces cannot be identified by ordinary people, and 50% of fake faces cannot be identified even by experienced photographers. Occasionally the neural network makes mistakes, producing artifacts such as incorrect curved patterns, strange hair colors, etc. Identifying fake faces requires careful observation. Common flaws include: symmetry issues (such as glasses, earrings), uneven teeth, repeated front teeth, hair with halos or too straight and striped, backgrounds with distortion or a torn feeling, bright colors bleeding from the background into the head, and the glossy "water stains" characteristic of the NVIDIA StyleGAN algorithm. Jevin West and Carl Bergstrom created the "Which Face Is Real" website to teach people to look more analytically at potentially fake portraits.
