mapping network

**Mapping network** is the **latent-transformation module that converts input noise vectors into intermediate latent representations optimized for style control** - it decouples sampling space from synthesis-control space. **What Is Mapping network?** - **Definition**: Typically an MLP that maps Z-space inputs to intermediate W-space embeddings. - **Functional Purpose**: Reshapes latent distribution to improve disentanglement and controllability. - **Architecture Position**: Sits between random latent sampling and generator style modulation layers. - **Output Usage**: Generated codes drive per-layer style parameters in synthesis network. **Why Mapping network Matters** - **Disentanglement Gains**: Improves separation of semantic factors compared with raw latent input. - **Editing Quality**: Enables smoother and more predictable latent manipulations. - **Training Stability**: Helps absorb latent-distribution irregularities before generation. - **Control Flexibility**: Supports truncation and style-mixing workflows in inference. - **Model Performance**: Contributes to higher fidelity and better latent-space geometry. **How It Is Used in Practice** - **Depth Selection**: Tune mapping-network layers to balance expressiveness and overfitting risk. - **Regularization**: Use path-length and style-mixing regularization to shape latent behavior. - **Latent Probing**: Evaluate semantic smoothness and attribute linearity in mapped space. Mapping network is **a key latent-conditioning component in modern style-based generators** - mapping-network design strongly affects editability and generative robustness.