stylegan
**StyleGAN (Style-based Generative Adversarial Network)** is a GAN architecture introduced by Karras et al. (2019) that generates high-fidelity images through a style-based generator design, where a learned mapping network transforms a latent code z into an intermediate latent space W, and adaptive instance normalization (AdaIN) injects these style vectors at each resolution level of the synthesis network. This design provides unprecedented control over generated image attributes at different spatial scales.
**Why StyleGAN Matters in AI/ML:**
StyleGAN set the **quality benchmark for unconditional image generation** and introduced the disentangled W latent space that enabled intuitive, hierarchical control over generated images from coarse structure to fine details, becoming the foundation for modern GAN-based generation and editing.
• **Mapping network** — An 8-layer MLP transforms the random latent z ∈ Z into an intermediate latent w ∈ W that is better disentangled than Z; the W space separates high-level attributes (pose, identity) from low-level details (hair texture, skin), enabling more meaningful interpolation
• **Adaptive Instance Normalization (AdaIN)** — Style vectors derived from w are injected at each generator layer via AdaIN: normalized features are scaled and shifted by learned affine transformations of w, providing per-layer control over the generated style
• **Hierarchical style control** — Styles injected at low resolutions (4×4-8×8) control coarse features (pose, face shape); mid-resolutions (16×16-32×32) control medium features (facial features, hairstyle); high resolutions (64×64+) control fine details (color, texture, microstructure)
• **Style mixing** — Using different w vectors at different layers (style mixing regularization) during training improves disentanglement and enables compositional generation: coarse structure from one image, fine details from another
• **Progressive improvements** — StyleGAN2 removed artifacts (water droplet artifacts from AdaIN, phase artifacts from progressive growing) with weight demodulation and skip connections; StyleGAN3 achieved alias-free generation with continuous signal processing
| Version | Key Innovation | Resolution | FID (FFHQ) |
|---------|---------------|-----------|------------|
| StyleGAN | Style-based synthesis, mapping network | 1024² | 4.40 |
| StyleGAN2 | Weight demodulation, no progressive | 1024² | 2.84 |
| StyleGAN2-ADA | Adaptive discriminator augmentation | 1024² | 2.42 |
| StyleGAN3 | Alias-free, continuous equivariance | 1024² | 4.40 (but alias-free) |
| StyleGAN-XL | Scaling to ImageNet | 1024² | 2.30 (ImageNet) |
**StyleGAN revolutionized image generation by introducing the style-based synthesis paradigm with its disentangled W latent space and hierarchical style injection, providing unprecedented control over generated image attributes at every spatial scale and establishing the architecture that defined the quality frontier for GAN-based image synthesis across multiple subsequent generations.**