Implementation A Style-Based Generator Architecture for Generative Adversarial Networks in PyTorch
Branch: master
Clone or download
Latest commit 192b1d1 Feb 11, 2019
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
checkpoint Fixed mini batch discrimination; DataParallel; PixelNorm on MLP Feb 8, 2019
doc Updated samples Feb 11, 2019
sample
.gitignore Initial commit Jan 29, 2019
LICENSE Initial commit Jan 29, 2019
README.md Updated samples Feb 11, 2019
generate.py Updated samples Feb 11, 2019
model.py Updated samples Feb 11, 2019
train.py Fixed mini batch discrimination; DataParallel; PixelNorm on MLP Feb 8, 2019

README.md

Style-Based GAN in PyTorch

Implementation of A Style-Based Generator Architecture for Generative Adversarial Networks (https://arxiv.org/abs/1812.04948) in PyTorch

Usage:

python train.py -d {celeba} PATH

Sample

Sample of the model trained on CelebA Style mixing sample of the model trained on CelebA

I have mixed styles at 4^2 - 8^2 scale. I can't get samples as dramatic as samles in the original paper. I think my model too dependent on 4^2 scale features - it seems like that much of details determined in that scale, so little variations can be acquired after it.