Progressive Growing/StyleGAN/StyleGAN2

转载自： https://www.notion.so/Progressive-Growing-StyleGAN-StyleGAN2-83859f6f47ba4a508375e4e4cf0efd05
教学视频： https://www.youtube.com/playlist?list=PLl6jkZVaP8zrWGBRTmcy_HeVfDq27oXML

参考文献：

• Progressive Growing of GANs for Improved Quality, Stability, and Variation
• A Style-Based Generator Architecture for Generative Adversarial Networks
• Analyzing and Improving the Image Quality of StyleGAN
• https://nvlabs.github.io/stylegan2/versions.html

Progressive Growing GAN

Picture: Two imaginary celebrities that were dreamed up by a random number generator.

Result Video

https://www.youtube.com/watch?v=G06dEcZ-QTg

More results

Main idea: Progressive Growth

The idea of multi resolution, multi-scale, pyramid, progressive itself is straightforward and not new.

For example LAPGAN back from 2015, SGAN 2016,

LAPGAN sampling

LAPGAN training

PGGAN

StyleGAN

AdaIN

In StyleGAN, $y$ is computed from $w$ instead of an example image.

Perceptual path length

Intend to measure how smooth the interpolation of latent vectors is.

The idea is to want the path length to be short in some perceptual space.

Truncation trick

StyleGAN2

https://www.youtube.com/watch?v=c-NJtV9Jvp0

Droplet-like artifacts→ alternative normalization

• modulation: scaling the convolution weights instead of the features: $w_{ijk}^{\prime }=s_i\cdot w_{ijk}$
• by assuming the input activations are i.i.d with unit standard deviation, after modulation an c convolution, the output have standard deviation of ${\sigma }_j=\sqrt{{\sum }_{i,j}{w_{ijk}^{\prime }}^2}$
• so, demodulation: $w_{ijk}^{\prime \prime }=\frac{w_{ijk}^{\prime }}{\sqrt{{\sum }_{i,k}{w_{ijk}^{\prime }}^2+ϵ}}$
• The style process can be baked into a convolution
• a.k.a weight normalization
• removes the artifacts, same FID.

Image quality and generator smoothness

The regularization

Can compute efficiently with

• $a$ is set dynamically during optimization as the long running exponential moving average of the lengths $\Vert {\mathbf{J}}_{\mathbf{w}}\mathbf{y}{\Vert }_2$.
• closely related to Jacobian clamping regularization.
• Better FID than spectral normalization

Progressive growing revisited