Home

Awesome

PyTorch Implementation of EigenGAN

Train

python train.py [image_folder_path] --name [experiment name]

Test

python test.py [ckpt path] --traverse

FFHQ

[ckpt]

samples (no truncation)

./docs/ffhq/sample.jpg

Learned subspace: L0 D1 ./docs/ffhq/traverse_L0_D1.jpg

Learned subspace: L1 D2 ./docs/ffhq/traverse_L1_D2.jpg

Anime

[ckpt]

samples (no truncation)

./docs/anime/sample.jpg

Learned subspace: L0 D5 ./docs/anime/traverse_L0_D5.jpg

Learned subspace: L1 D2 ./docs/anime/traverse_L1_D2.jpg

Note

Default training configurations are different from the original implementation

Tested on python 3.8 + torch 1.8.1

Issue

Some of the subspace layers seem to collapse (does not contribute to the output) as the training proceeds and FID gets better.