Mapping Network

The mapping network's primary function is to transform a latent vector (z) sampled from a simple distribution (e.g., Gaussian) into an intermediate latent space (W). This is achieved via an 8-layer feed-forward neural network. The transformation non-linearly warps the input space, disentangling entangled factors of variation present in z. This creates a W-space where directions correspond to more interpretable and independent image attributes, such as pose, lighting, and hairstyle.

A core achievement of the mapping network is learning disentangled representations. By projecting into the intermediate W-space, the network organizes semantic features along orthogonal axes. This means a single dimension in W can control a specific attribute (e.g., age) with minimal effect on others (e.g., identity). This disentanglement is what enables the style mixing technique, where styles from two different W vectors can be applied to different layers of the synthesis network to combine features seamlessly.

The mapping network is typically implemented as a multi-layer perceptron (MLP). Its architecture is designed to apply a series of non-linear transformations:

• Layers: Commonly 8 fully-connected layers.
• Activation: Each layer uses a leaky ReLU activation function.
• Normalization: Features are normalized after each layer, often using a form of layer normalization. This deep, non-linear processing is essential for breaking the correlations present in the input Gaussian noise and mapping it to a space that the synthesis network can interpret as independent style parameters.

The output of the mapping network—the W vector—does not directly generate pixels. Instead, it serves as the style input to the synthesis network. For each convolutional layer in the synthesis network, the W vector is transformed by a separate, learned affine transformation (A) to create a style vector (y). This style vector modulates the convolutional feature maps via Adaptive Instance Normalization (AdaIN), aligning their channel-wise statistics (mean and variance) with the style. This allows hierarchical control, with coarse styles (from earlier W inputs) affecting high-level structure and fine styles affecting details.

In a traditional GAN, the generator receives a latent vector z directly. This often leads to entangled representations, where changing one dimension affects multiple attributes unpredictably. The mapping network introduces a key abstraction:

• Traditional GAN: Latent code z → Generator → Image.
• StyleGAN with Mapping Network: Latent code z → Mapping Network (f) → Intermediate code w → Synthesis Network → Image. This decoupling allows the synthesis network to operate on a learned, structured space (W) that is better suited for generating coherent and controllable images than the raw input noise.

The mapping network contributes significantly to the training stability and final output quality of StyleGAN. By providing a learned, normalized input space for the synthesis network, it reduces the complexity the generator must handle in a single step. This separation of concerns—where the mapping network learns the distribution of styles and the synthesis network learns the rendering process—mitigates common GAN failure modes like mode collapse. It also enables advanced techniques like the truncation trick, where sampling W vectors closer to the average can boost fidelity at the cost of diversity.

The synthesis network is the primary image-generating component in StyleGAN. It is a convolutional neural network that starts from a learned constant input tensor and progressively upsamples it to the final output resolution. Crucially, its layers are modulated by style vectors (w) produced by the mapping network via Adaptive Instance Normalization (AdaIN), allowing the mapping network's output to control visual attributes at different hierarchical levels (coarse styles like pose at early layers, fine details like hair color at later layers).

Adaptive Instance Normalization (AdaIN) is the critical operation that connects the mapping network to the synthesis network. For each convolutional feature map in the synthesis network, AdaIN applies the following transformation:

• It first normalizes the feature map to have zero mean and unit variance (instance normalization).
• It then scales and shifts these normalized features using the affine parameters (scale and bias) derived from the current style vector (w). This mechanism allows a single style vector from the mapping network to control the statistical properties (style) of the generated image at a specific layer, enabling precise, disentangled manipulation of image features.

StyleGAN operates across two distinct latent spaces:

• Z-space: The input Gaussian latent space. This is a 512-dimensional space where random vectors (z) are sampled. It typically follows a standard normal distribution and exhibits entangled representations.
• W-space (Intermediate Latent Space): The output space of the mapping network. The mapping network's primary function is to transform a z-vector into a w-vector in this space. W-space is empirically found to be more linearly separable and disentangled, meaning directions in this space correspond to more interpretable and independent image attributes (e.g., changing age without affecting pose). This disentanglement is what enables high-quality style mixing and editing.

Style mixing is a demonstration of the disentanglement achieved by the mapping network and AdaIN. During generation, instead of using a single w-vector for all layers of the synthesis network, different w-vectors (from different source latents) can be used for different coarse, middle, and fine layers. For example:

• Coarse styles (resolution 4x4 to 8x8): Control high-level attributes like pose, face shape, and hairstyle.
• Middle styles (16x16 to 32x32): Control facial features and eye gaze.
• Fine styles (64x64 to 1024x1024): Control color scheme and micro-details. This is possible because the mapping network produces a w-vector that the synthesis network can apply per-layer, allowing for independent control over these hierarchical aspects.

Feature disentanglement is the desired property where distinct semantic attributes of the generated data are controlled by separate and independent factors in the latent space. The mapping network is engineered to promote this in W-space. Unlike the entangled Z-space, traversing along a single axis in W-space (after mapping) tends to vary a single interpretable attribute (e.g., smile intensity, yaw rotation) while leaving others largely unchanged. This is measured via linear separability tests and perceptual path length metrics. Disentanglement is crucial for reliable, predictable image editing applications.

The truncation trick is a post-processing technique applied to the latent vectors fed into the mapping network. Instead of sampling the input z-vector from the full normal distribution N(0, I), it is sampled from a truncated distribution (e.g., N(0, I) with values clipped beyond a threshold ψ). This pulls the latent vectors closer to the distribution's mean. When these truncated z-vectors are passed through the mapping network, the resulting w-vectors tend to produce higher fidelity, more "average" looking images but at the cost of reduced diversity. It provides a practical trade-off knob between quality and variety in the final output.