rosinality / stylegan2-pytorch

Implementation of Analyzing and Improving the Image Quality of StyleGAN (StyleGAN 2) in PyTorch

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

StyleGAN 2 in PyTorch

Implementation of Analyzing and Improving the Image Quality of StyleGAN ( in PyTorch


I have tried to match official implementation as close as possible, but maybe there are some details I missed. So please use this implementation with care.


I have tested on:

  • PyTorch 1.3.1
  • CUDA 10.1/10.2


First create lmdb datasets:

python --out LMDB_PATH --n_worker N_WORKER --size SIZE1,SIZE2,SIZE3,... DATASET_PATH

This will convert images to jpeg and pre-resizes it. This implementation does not use progressive growing, but you can create multiple resolution datasets using size arguments with comma separated lists, for the cases that you want to try another resolutions later.

Then you can train model in distributed settings

python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT --batch BATCH_SIZE LMDB_PATH supports Weights & Biases logging. If you want to use it, add --wandb arguments to the script.


This implementation experimentally supports SWAGAN: A Style-based Wavelet-driven Generative Model ( You can train SWAGAN by using

python -m torch.distributed.launch --nproc_per_node=N_GPU --master_port=PORT --arch swagan --batch BATCH_SIZE LMDB_PATH

As noted in the paper, SWAGAN trains much faster. (About ~2x at 256px.)

Convert weight from official checkpoints

You need to clone official repositories, ( as it is requires for load official checkpoints.

For example, if you cloned repositories in ~/stylegan2 and downloaded stylegan2-ffhq-config-f.pkl, You can convert it like this:

python --repo ~/stylegan2 stylegan2-ffhq-config-f.pkl

This will create converted file.

Generate samples

python --sample N_FACES --pics N_PICS --ckpt PATH_CHECKPOINT

You should change your size (--size 256 for example) if you train with another dimension.

Project images to latent spaces


Closed-Form Factorization (

You can use and to discover meaningful latent semantic factor or directions in unsupervised manner.

First, you need to extract eigenvectors of weight matrices using


This will create factor file that contains eigenvectors. (Default: And you can use to test the meaning of extracted directions


For example,

python -i 19 -d 5 -n 10 --ckpt [CHECKPOINT]

Will generate 10 random samples, and samples generated from latents that moved along 19th eigenvector with size/degree +-5.

Sample of closed form factorization

Pretrained Checkpoints


I have trained the 256px model on FFHQ 550k iterations. I got FID about 4.5. Maybe data preprocessing, resolution, training loop could made this difference, but currently I don't know the exact reason of FID differences.


Sample with truncation

Sample from FFHQ. At 110,000 iterations. (trained on 3.52M images)

MetFaces sample with non-leaking augmentations

Sample from MetFaces with Non-leaking augmentations. At 150,000 iterations. (trained on 4.8M images)

Samples from converted weights

Sample from FFHQ

Sample from FFHQ (1024px)

Sample from LSUN Church

Sample from LSUN Church (256px)


Model details and custom CUDA kernel codes are from official repostiories:

Codes for Learned Perceptual Image Patch Similarity, LPIPS came from

To match FID scores more closely to tensorflow official implementations, I have used FID Inception V3 implementations in


Implementation of Analyzing and Improving the Image Quality of StyleGAN (StyleGAN 2) in PyTorch

License:MIT License


Language:Python 90.2%Language:Cuda 8.3%Language:C++ 1.5%