Home▸Machine Learning & Neural Networks▸Style-Based Image Generation: Controlling GANs at Every Layer

🎭 Style-Based Image Generation: Controlling GANs at Every Layer

Explore how style-based GANs map random noise into a style space and inject it at every resolution, giving disentangled, layer-by-layer control over generated images.

Machine Learning & Neural Networks3DModerate60 FPS

A 3D visualisation of a style-based generator pipeline, showing a latent vector passing through a mapping network into a style space that is then injected into successive resolution layers of a synthesis network.

🔬 What It Demonstrates

🎮 How to Use

Select which layer group (coarse, middle or fine) receives an alternate style vector, then use the speed slider and play/pause to animate how that style mixing propagates through the network and reshapes the output.

💡 Did You Know?

In the original StyleGAN paper, mixing styles from two different latent codes during training acted as a regulariser that prevented the generator from assuming neighbouring layers always share correlated styles.

⚙ Under the hood

Explore how coarse, middle and fine layers of a style-based generator each control a different level of visual detail, from pose down to fine texture.

styleganganimage generationstyle transfermachine-learninggenerative models

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install