Chainer implementation of Style-based Generator

A Style-Based Generator Architecture for Generative Adversarial Networks

https://arxiv.org/abs/1812.04948

Requirements

opencv-python
python-gflags
Augmentor
h5py
Pillow
scipy
mpi4py
chainer >= 5.0.0
cupy >= 5.0.0

• python >= 3.6.0
• NVIDIA driver 391.35 or newer, CUDA toolkit 9.0 or newer, cuDNN 7.3.1 or newer.
• NCCL2
• A graphic card with at least 11GB memory to train the 1024x1024 model.
• Tested on 8 Tesla P100.

Datasets

1. Please follow ffhq-dataset to obtain the ffhq dataset.

python download_ffhq.py -h -i

2. Convert raw ffhq images to a HDF5 file. (Around 198GB)

cd src/hdf5_tools
bash folder_to_multisize_hdf5_cmds.sh 1 YOUR_PATH_TO_RAW_FFHQ_IMAGES

Run

• 8 GPUs setting

cd src/stylegan
bash run_ffhq.sh 2 

• 1 GPU setting (up to 256x256)

cd src/stylegan
bash run_ffhq.sh 1

Pre-trained Model on FFHQ

GDrive

Sampling images on CPU python sampling.py --m_style SmoothedGenerator_405000.npz --m_mapping SmoothedMapping_405000.npz --gpu -1

Results