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

Datasets

  1. Please follow ffhq-dataset to obtain the ffhq dataset.
python download_ffhq.py -h -i
  1. 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

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

samples