StyleGAN: Style-Based Generator Architecture
What you'll learn: How StyleGAN revolutionizes image generation by giving unprecedented control over style at multiple scales through mapping networks and adaptive instance normalization.
The Core Innovation
Traditional GANs feed random noise directly into the generator. StyleGAN takes a radically different approach: it separates the concepts of content and style, then injects style information at multiple resolution levels during generation.
Think of it like a professional photographer working with a model. The photographer doesn't control everything at once—they adjust lighting (overall mood), makeup (fine details), and pose (structure) independently. StyleGAN works similarly, controlling different aspects of an image separately.
Three Key Components
Mapping Network
Instead of feeding noise z directly into the generator, StyleGAN first passes it through a mapping network (typically 8 fully-connected layers). This transforms z into an intermediate latent space W, which is more disentangled—meaning individual dimensions control specific, interpretable features (like age, pose, or lighting).
Adaptive Instance Normalization (AdaIN)
At each resolution level in the generator, AdaIN takes the intermediate representation and "styles" it using learned parameters from W. It normalizes the feature maps, then scales and shifts them based on the style vector—like applying a filter that says "make this look more photorealistic" or "add this lighting characteristic."
Progressive Generation
StyleGAN generates images progressively from coarse to fine resolutions (building on ProGAN). Early layers control high-level aspects (pose, general structure), while later layers control fine details (hair texture, skin pores).
Key Takeaway: StyleGAN's mapping network, AdaIN layers, and progressive architecture enable unmatched control over image synthesis by disentangling and injecting style information at multiple scales, making specific attributes independently modifiable without affecting others.