Progressive Growing of GANs

The paper made high-resolution GAN training practical by growing the generator and discriminator from coarse to fine resolutions.

Metadata

Field	Value
Paper	Progressive Growing of GANs for Improved Quality, Stability, and Variation
Authors	Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen
Year	2017 preprint; 2018 conference
Venue	ICLR 2018
arXiv	1710.10196
OpenReview	Hk99zCeAb
Code	tkarras/progressive_growing_of_gans
Status	verified

Question

DCGAN gave a practical convolutional image-GAN recipe, but directly training high-resolution GANs remains unstable. The generator and discriminator must learn global structure and fine detail at the same time.

Progressive GAN asks:

$$\text{Can high-resolution GAN training be stabilized by learning coarse structure first?}$$

The answer is to start at low resolution and add layers gradually.

Main Claim

The paper grows both networks over training:

$$4^2\to 8^2\to 16^2\to \cdots \to 1024^2.$$

The durable architecture/training claim is:

$$\text{progressive generator growth}+\text{progressive discriminator growth}+\text{fade-in transitions}\Rightarrow \text{more stable high-resolution GAN synthesis}.$$

This belongs in architecture papers because it changes the network topology during training.

Architecture Contract

Item	Contract
Latent input	$z\sim p(z)$
Generator	starts from low-resolution block and adds higher-resolution blocks
Discriminator	mirrors the current image resolution and adds matching blocks
Training schedule	alternate stabilization and fade-in phases
Output resolution	increases during training
Core mechanism	fade-in interpolation between old and new paths
Main target	high-resolution image generation

Progressive Growth

At resolution $r$, the generator produces:

$${\hat{x}}^{(r)}=G^{(r)}(z).$$

Training starts with a small network:

$$G^{(4)}:z\to 4\times 4.$$

Then new blocks are appended:

$$G^{(8)}={\mathrm{ToRGB}}_8\circ B_8\circ G^{(4)}.$$

The discriminator grows symmetrically:

$$D^{(8)}=D^{(4)}\circ B_8^D\circ {\mathrm{FromRGB}}_8.$$

The model therefore learns:

Stage	Main Burden
low resolution	global layout and coarse object structure
middle resolution	parts and medium-scale consistency
high resolution	texture and fine detail

Fade-In Transition

When adding a new resolution, the output blends old and new paths:

$$x_{\text{out}}=(1-\alpha )x_{\text{old}}+\alpha x_{\text{new}},\alpha \in [0,1].$$

As training proceeds:

$$\alpha :0\to 1.$$

This avoids abruptly changing the generator/discriminator interface.

The same idea applies on the discriminator side, where real and generated images at the new resolution are blended into the old path during transition.

Why It Helps

Direct high-resolution GAN training couples too many tasks:

$$\text{layout}+\text{shape}+\text{texture}+\text{local detail}\text{all at once}.$$

Progressive growing separates them over training time:

$$\text{coarse}\to \text{medium}\to \text{fine}.$$

This is not only a curriculum over data resolution. It is a curriculum over model capacity and topology.

Practical Architecture Details

The paper also uses implementation choices that became associated with the Progressive GAN/StyleGAN lineage.

Detail	Role
minibatch standard deviation	encourage sample variation
equalized learning rate	normalize dynamic range of weights during training
pixelwise feature normalization	control feature magnitude in generator
fade-in layers	avoid sudden topology changes
progressive dataset resolution	match model resolution during training

These details matter because GAN training stability depends on the interaction between generator and discriminator.

Relation to StyleGAN

StyleGAN inherits the high-resolution synthesis lineage but changes generator control:

Paper	Main Idea
Progressive GAN	grow networks from low to high resolution
StyleGAN	style-based per-layer modulation and stochastic detail control

Progressive GAN is the training-topology milestone before style-based generator control.

Evidence

Evidence Type	What It Supports
high-resolution face synthesis	progressive growing can reach $1024^2$ image generation
stability observations	growing topology eases adversarial training
variation metric	evaluates both quality and diversity
CIFAR/ImageNet/CelebA-HQ experiments	method generalizes beyond one visual dataset

Limits

• Progressive growing is a training procedure, not a universal generator block.
• Fade-in schedules add implementation complexity.
• Later StyleGAN variants moved away from progressive growing in some settings.
• High visual quality can still depend heavily on dataset curation and compute.

Concepts

• GAN
• CNN
• Sampling
• Generation evaluation

Related

• DCGAN
• Generative Adversarial Nets
• StyleGAN
• Self-Attention GAN