Key points
- StyleGAN injects style at each layer, giving disentangled control over generated faces and characters.
- It set the standard for synthetic face and character generation in animation, games and digital-human work.
- It spawned StyleGAN2/3 and a large latent-editing family, and made AI faces culturally prominent.
Summary
Karras, Laine and Aila (2019) redesigned the GAN generator to inject learned style vectors at each resolution level via adaptive instance normalisation, replacing the traditional input latent with a mapping network that produces a disentangled intermediate latent space. The architecture gives fine-grained independent control over coarse features such as pose and face shape, mid-level features such as hairstyle, and fine-detail features such as skin texture. StyleGAN set the benchmark for photorealistic synthetic face generation and became the reference architecture for generative character work in animation, games and digital-human pipelines. Its successors (StyleGAN2, StyleGAN3) and the latent-editing methods built on its intermediate space form a widely taught family in AI-art and character-generation courses.
Related items
Source
Source: CVPR 2019 (IEEE/CVF) ↗ (Research)
Cite this item
Karras, T., Laine, S. & Aila, T. (2019). ‘A Style-Based Generator Architecture for GANs (StyleGAN)’, CVPR 2019 (IEEE/CVF). Available at: https://arxiv.org/abs/1812.04948
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.