I found bugs in the implementation thanks to @adambielski and @TropComplique! (rosinality#33, rosinality#34) I have fixed this and updated checkpoints
- Now trainer uses pre-resized lmdb dataset for more stable data loading and training.
- Model architecture is now more closely matches with official implementation.
Implementation of A Style-Based Generator Architecture for Generative Adversarial Networks (https://arxiv.org/abs/1812.04948) in PyTorch
Usage:
You should prepare lmdb dataset
python prepare_data.py --out LMDB_PATH --n_worker N_WORKER DATASET_PATH
This will convert images to jpeg and pre-resizes it. (For example, 8/16/32/64/128/256/512/1024) Then you can train StyleGAN.
for celebA
python train.py --mixing LMDB_PATH
for FFHQ
python train.py --mixing --loss r1 --sched LMDB_PATH
| Resolution | Model & Optimizer |
|---|---|
| 256px | Link |
| 512px | Link |
| 1024px | Link |
Model & Optimizer checkpoints saved at the end of phases of each resolution. (that is, 512px checkpoint saved at the end of 512px training.)
512px sample from the generator trained on FFHQ.
| Resolution | Model & Optimizer | Running average of generator |
|---|---|---|
| 128px | Link | 100k iter Link |
| 256px | Link | 140k iter Link |
| 512px | Link | 180k iter Link |
Old version of checkpoints. As gradient penalty and discriminator activations are different, it is better to use new checkpoints to do some training. But you can use these checkpoints to make samples as generator architecture is not changed.
Running average of generator is saved at the specified iterations. So these two are saved at different iterations. (Yes, this is my mistake.)