StyleGAN2

The paper redesigned StyleGAN’s normalization and regularization to reduce artifacts and improve image quality.

Metadata

Field	Value
Paper	Analyzing and Improving the Image Quality of StyleGAN
Authors	Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, Timo Aila
Year	2019 preprint; 2020 conference
Venue	CVPR 2020
arXiv	1912.04958
CVF	CVPR 2020 paper
Code	NVlabs/stylegan2
Status	verified

Question

StyleGAN made style-based synthesis a canonical high-resolution GAN architecture, but it also introduced characteristic artifacts. StyleGAN2 asks:

$$\text{Which parts of StyleGAN’s generator architecture cause artifacts, and how should they be redesigned?}$$

The answer is a revised generator with weight modulation/demodulation, no progressive growing, and path length regularization.

Main Claim

StyleGAN2 keeps the style-based synthesis idea but changes how styles affect convolutional weights and how the generator is regularized.

The durable architecture claim is:

$$\text{style modulation}+\text{weight demodulation}+\text{skip/residual generator redesign}+\text{path length regularization}\Rightarrow \text{cleaner high-resolution GAN synthesis}.$$

This is not just a training tweak. It is a generator-architecture correction paper.

Architecture Contract

Item	Contract
Input	latent code $z$ mapped to intermediate style $w$
Generator	style-based convolutional synthesis network
Main change	replace AdaIN-style feature normalization with weight modulation and demodulation
Training topology	avoid progressive growing in favor of a fixed architecture
Regularization	path length regularization for smoother latent-to-image mapping
Output	high-resolution image
Main target	reduce StyleGAN artifacts and improve image quality

From AdaIN to Weight Demodulation

StyleGAN injects style through adaptive instance normalization:

$$\mathrm{AdaIN}(x,y)=y_s\frac{x-\mu (x)}{\sigma (x)}+y_b.$$

StyleGAN2 argues that this feature normalization can contribute to artifacts. The redesigned view modulates convolution weights directly.

Let $w$ be a style vector and let $s_i$ be a per-input-channel style scale. For convolution weight:

$$W_{ijk},$$

where $i$ indexes output channel, $j$ input channel, and $k$ spatial kernel position, modulation gives:

$$W_{ijk}^{\prime }=s_j{W}_{ijk}.$$

Demodulation normalizes each output channel:

$$W_{ijk}^{\prime \prime }=\frac{W_{ijk}^{\prime }}{\sqrt{{\sum }_{j,k}(W_{ijk}^{\prime }{)}^2+ϵ}}.$$

This preserves style control while reducing signal magnitude artifacts.

Why Remove Progressive Growing?

Progressive GAN grows the network resolution during training:

$$4^2\to 8^2\to \cdots \to 1024^2.$$

StyleGAN inherited this lineage. StyleGAN2 revisits it because progressive growing can leave resolution-dependent artifacts and makes the generator/discriminator interface change during training.

The revised approach uses a fixed generator architecture:

$$\text{all resolutions exist from the start}.$$

This makes training topology more stable and easier to analyze.

Path Length Regularization

The mapping from latent $w$ to image should have a well-conditioned Jacobian. StyleGAN2 encourages small and consistent changes in image space for changes in latent space.

Let:

$$G(w)\in {\mathbb{R}}^{H\times W\times C}.$$

The local sensitivity is related to:

$$J_w=\frac{\partial G(w)}{\partial w}.$$

Path length regularization penalizes deviations in the norm of image-space changes induced by latent perturbations:

$${\mathcal{L}}_{\text{pl}}={\mathbb{E}}_{w,y}{\left({\Vert J_w^{\top }y\Vert }_2-a\right)}^2,$$

where $y$ is a random image-space direction and $a$ is a moving average target length.

The practical effect is smoother latent interpolation and easier generator inversion.

Generator Redesign

StyleGAN2 compares generator structures and uses skip or residual paths rather than relying on progressive growing.

Component	StyleGAN	StyleGAN2
style injection	AdaIN on activations	weight modulation/demodulation
training topology	progressive growing	fixed architecture
artifact diagnosis	observed after synthesis	explicitly analyzed and targeted
latent smoothness	style mixing and metrics	path length regularization

Relation to StyleGAN Lineage

Paper	Main Contribution
Progressive GAN	grow generator/discriminator resolution over training
StyleGAN	style-based per-layer synthesis control
StyleGAN2	weight demodulation and fixed generator redesign to reduce artifacts

The sequence is useful because it shows how high-resolution GANs moved from training curriculum to generator control to artifact-aware architecture design.

Evidence

Evidence Type	What It Supports
artifact analysis	StyleGAN artifacts are tied to architecture/normalization behavior
image quality metrics	revised generator improves distribution quality metrics
perceptual comparisons	artifacts are reduced in generated images
path length regularization analysis	latent interpolation and inversion behavior improve

Limits

• The paper is centered on unconditional image generation.
• It improves the StyleGAN lineage but does not remove adversarial training tradeoffs.
• The architecture is still compute- and data-sensitive.
• Later StyleGAN variants address aliasing, data efficiency, and domain adaptation separately.

Concepts

• GAN
• Convolution
• Normalization
• Sampling
• Generation evaluation

Related

• StyleGAN
• Progressive Growing of GANs
• BigGAN
• Generative Adversarial Nets