← 🎭 Machine Learning & Neural Networks

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

Active layer:
Style mix: 0%
Drag — rotate · Scroll — zoom

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

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

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.

🎮 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.