A Style-Based Generator Architecture for Generative Adversarial Networks

Problem

Framing

Progressive GANs generate sharp high-resolution images, but a single input latent leaves semantics entangled across layers and scales. StyleGAN closes this gap with an intermediate latent $\mathbf{w}$, per-layer style control, and explicit noise inputs, reducing FFHQ FID from 8.04 to 4.40.

Currently Used Methods

Foundational

• @goodfellowGAN2014 — adversarial training for implicit generative modeling.
  • Limitation in context: no mechanism for scale-specific latent control.
• @radfordDCGAN2015 — convolutional GAN design for image synthesis.
  • Limitation in context: latent factors stay mixed through all generator layers.
• Progressive Growing of GANs for Improved Quality, Stability, and Variation — stable high-resolution GAN training by progressive layer growth.
  • Limitation in context: generator semantics remain entangled across resolutions.
• Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization — AdaIN modulates features with channelwise affine parameters.
  • Limitation in context: not a generative latent architecture with learned styles.

Proposed Method

Architecture

The generator maps $\mathbf{z} \in \mathcal{Z}$ through an 8-layer MLP into $\mathbf{w} \in \mathcal{W}$. A separate 18-layer synthesis network starts from a learned $4 \times 4 \times 512$ constant, applies AdaIN after each convolution, and injects single-channel Gaussian noise at every layer.

Loss / Objective

The paper keeps the GAN objective and changes the generator parameterization through adaptive instance normalization.

Sampling Rule / Algorithm

Sampling maps $\mathbf{z}$ into layerwise styles, then synthesizes from a learned constant plus stochastic noise.

Training Procedure

• Mapping network depth: 8 fully connected layers.
• Intermediate latent dimensionality: 512.
• Synthesis network: 18 layers.
• Learned input constant: $4 \times 4 \times 512$.
• Mixing regularization tested at 0%, 50%, 90%, 100%.
• CelebA-HQ loss: WGAN-GP.
• FFHQ loss: non-saturating logistic with R1 regularization.

Evaluation

Datasets

• CelebA-HQ
• FFHQ
• LSUN Bedroom
• LSUN Car

Metrics

• FID
• Perceptual path length
• Linear separability

Headline results

• CelebA-HQ: baseline 7.79 FID; full model 5.06 FID.
• FFHQ: baseline 8.04 FID; full model 4.40 FID.
• FFHQ tuned baseline: 5.25 FID.
• LSUN Bedroom at $256^2$: 2.65 FID.
• LSUN Car at $512 \times 384$: 3.27 FID.

Table 1: FID for generator variants on CelebA-HQ and FFHQ

Ablations

• Mapping network plus styles sharply improves FID over the tuned baseline.
• Replacing the latent input with a learned constant helps CelebA-HQ more than FFHQ.
• Noise inputs improve CelebA-HQ FID from 5.07 to 5.06.
• Mixing regularization gives the best FFHQ FID and improves mixed-latent robustness.

Method Strengths and Weaknesses

Strengths

• Intermediate $\mathcal{W}$ improves separability over direct $\mathcal{Z}$ input.
• Per-layer styles expose coarse-to-fine semantic control.
• Noise inputs isolate stochastic detail from global structure.
• Full design cuts FFHQ FID from 8.04 to 4.40.

Weaknesses

• Mixing regularization slightly worsens CelebA-HQ FID from 5.06 to 5.17.
• Gains come from several coupled changes, not one clean intervention.
• Evaluation centers on faces and two LSUN domains.
• The method still depends on GAN optimization stability.

Suggestions from the authors

• Analyze why the intermediate latent space improves disentanglement.
• Study truncation tradeoffs between fidelity, variation, and coverage.
• Characterize style control across coarse, middle, and fine scales.
• Extend the architecture to broader image domains.

Links

Prior Papers

• @goodfellowGAN2014 — foundational GAN objective retained while the generator parameterization is redesigned.
• @radfordDCGAN2015 — early convolutional GAN baseline that StyleGAN departs from architecturally.

Further Papers

No vault papers identified as further work yet.