Implementation of the paper ``Large Scale Image Completion via Co-Modulated Generative Adversarial Networks"
official tensorflow version: https://github.com/zsyzzsoft/co-mod-gan
conda install pytorch torchvision cudatoolkit=11 -c pytorch
conda install matplotlib jinja2 ninja dill
pip install git+https://github.com/zengxianyu/pytorch-fid
Download the code:
git clone https://github.com/zengxianyu/co-mod-gan-pytorch
git checkout train
git submodule init
git submodule update
- download pretrained model using ``download/*.sh" (converted from the tensorflow pretrained model)
e.g. ffhq512
./download/ffhq512.sh
converted model:
- FFHQ 512 checkpoints/comod-ffhq-512/co-mod-gan-ffhq-9-025000_net_G_ema.pth
- FFHQ 1024 checkpoints/comod-ffhq-1024/co-mod-gan-ffhq-10-025000_net_G_ema.pth
- Places 512 checkpoints/comod-places-512/co-mod-gan-places2-050000_net_G_ema.pth
- use the following command as a minimal example of usage
./test.sh
or
python test.py \
--how_many 10 \
--mixing 0 \
--batchSize 1 \
--nThreads 2 \
--name comod-ffhq-512 \
--dataset_mode testimage \
--image_dir ./ffhq_debug/images \
--mask_dir ./ffhq_debug/masks \
--output_dir ./ffhq_debug \
--load_size 512 \
--crop_size 512 \
--z_dim 512 \
--model comod \
--netG comodgan \
--which_epoch co-mod-gan-ffhq-9-025000
- download example datasets for training and validation
./download/data.sh
- use the following command as a minimal example of usage
./train.sh
Coming soon
[1] official tensorflow version: https://github.com/zsyzzsoft/co-mod-gan
[2] stylegan2-pytorch https://github.com/rosinality/stylegan2-pytorch
[3] pix2pixHD https://github.com/NVIDIA/pix2pixHD
[4] SPADE https://github.com/NVlabs/SPADE