1 of 14

reviews: StyleGAN

Abhinav Venigalla

2 of 14

history

Progressive Growing of GANs

StyleGAN

DC-GAN, BEGAN, etc.

Model:

3 of 14

history

Progressive Growing of GANs

StyleGAN

DC-GAN, BEGAN, etc.

CelebA

CelebA-HQ

FFHQ

Dataset:

Model:

4 of 14

history

Progressive Growing of GANs

StyleGAN

DC-GAN, BEGAN, etc.

CelebA

CelebA-HQ

FFHQ

Dataset:

Model:

NVIDIA Finland

(Tero Karras + friends)

5 of 14

architecture overview

​

  • progressively growing layers
    • 4x4 -> 8x8 -> … 1024x1024

​

  • z code sampled from 512-d Gaussian
  • w code is a function of z
  • w is fed into each layer as a style
    • AdaIn(xi ; f(w)))

​

  • noise is also added to each layer
    • to help generator produce stochastic outputs

​

​

6 of 14

7 of 14

latent code (z) -> style code (w)

  • Intuition: Sampling from high-d Gaussians is weird
    • “Gaussians are soap bubbles”
  • Mapping to a space where latent space is

disentangled would help

8 of 14

style mixing regularization

w_1

w_1

w_2

w_2

9 of 14

style mixing regularization

w_1

w_1

w_2

w_2

“prevents the network

from assuming

that adjacent styles

are correlated”

10 of 14

intentionally feeding noise as inputs

11 of 14

spin-off projects

  • Inverse-Z finding
    • Find a Z for a real person’s face
    • Gradient descent just works
    • Metropolis-Hastings sorta works

​

  • Face interpolations, style transfer

​

​

​

​

12 of 14

overall thoughts

  • progressive growing (still) works!
  • NO REGULARIZATION

​

​

​

​

​

13 of 14

overall thoughts

  • progressive growing (still) works!
  • NO REGULARIZATION

​

  • sending the latent code to each intermediate stage is good
  • having access to noise at intermediate stages helps the generator network
  • disentanglement of z -> w gives us better interpolations

​

​

​

14 of 14

overall thoughts

  • progressive growing (still) works!
  • NO REGULARIZATION

​

  • sending the latent code to each intermediate stage is good
  • having access to noise at intermediate stages helps the generator network
  • disentanglement of z -> w gives us better interpolations

​

  • computer graphics PHDs are op

​

​