Skip to content
Last Dictionary
Plane
Generative MediaDeep LearningSociety, Ethics & Economy

StyleGAN

proper noun/STYLE-gan/Also: StyleGAN2, StyleGAN2-ADA, StyleGAN3, style-based generator

In plain wordsAn image-generating network from NVIDIA, famous for photorealistic faces of people who do not exist, that sets coarse traits such as head angle and face shape separately from fine details such as coloring and skin texture.

  1. A generative adversarial networkgenerative adversarial network, n.: Two neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people. architecture whose generator injects a learned 'style' at every resolution. A mapping network of eight fully connected layers transforms the input noise vector z into an intermediate latent vector w; learned affine transforms of w then modulate each convolutional block of a synthesis network that starts from a learned constant tensor, originally through adaptive instance normalization (AdaIN), while per-layer noise inputs add stochastic detail such as the placement of hairs. Because w enters at every scale, styles can be mixed — coarse layers set pose and face shape, fine layers color and texture — and the intermediate W space is far more disentangled than z, making it a standard space for GAN inversion and semantic editing. StyleGAN2 (2020) replaced AdaIN with weight demodulation to remove blob-like artifacts, StyleGAN2-ADA added adaptive data augmentation for small datasets, and StyleGAN3 (2021) made the generator alias-free so that textures no longer stick to pixel coordinates.

    We projected the photo into StyleGAN's W space and nudged the 'age' direction.

Origin

Introduced by Tero Karras, Samuli Laine and Timo Aila of NVIDIA in a paper posted in December 2018 (CVPR 2019), together with FFHQ, a dataset of 70,000 high-quality face photographs from Flickr. Code followed in February 2019, when Phillip Wang's website This Person Does Not Exist, showing a new generated face on every reload, brought the model wide public attention.

Usage note

StyleGAN faces quickly became a staple of fake social-media profiles, an early mass use of synthetic mediasynthetic media, n.: Images, video, audio or text created or substantially altered by AI rather than recorded or written directly by people, whether for art, entertainment, business or deception.; see deepfakedeepfake, n.: A fake but realistic video, photo or voice recording made with AI that puts words in a real person's mouth or shows them in a scene that never happened..

Related

See also
  • generative adversarial networkgenerative adversarial network, n.: Two neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people.
  • style transferstyle transfer, n.: Repainting a photo in the style of an artwork — your street as Van Gogh might have painted it — by keeping the photo's content and borrowing the painting's colors, strokes and textures.
  • latent spacelatent space, n.: The hidden internal map a model builds of its data, where each item becomes a point and similar items sit near each other. Moving smoothly through it gradually changes what the model represents or generates.
  • deepfakedeepfake, n.: A fake but realistic video, photo or voice recording made with AI that puts words in a real person's mouth or shows them in a scene that never happened.
  • synthetic mediasynthetic media, n.: Images, video, audio or text created or substantially altered by AI rather than recorded or written directly by people, whether for art, entertainment, business or deception.
  • image-to-image translationimage-to-image translation, n.: Turning one picture into another version of itself — a sketch into a photo, a day scene into night, a horse into a zebra — while keeping its layout.

Sources

  1. Karras, T., Laine, S. & Aila, T. (2019). A Style-Based Generator Architecture for Generative Adversarial Networks. CVPR 2019. arXiv:1812.04948. arxiv.org/abs/1812.04948
  2. NVIDIA (2019). Flickr-Faces-HQ Dataset (FFHQ). GitHub. github.com/NVlabs/ffhq-dataset

Mentioned in

generative adversarial networkgenerative adversarial network, n.: Two neural networks trained as rivals: one forges fake images, the other tries to spot the fakes. As both improve, the forger's output becomes realistic enough to fool people., style transferstyle transfer, n.: Repainting a photo in the style of an artwork — your street as Van Gogh might have painted it — by keeping the photo's content and borrowing the painting's colors, strokes and textures.