OneshotCLIP

Official Source code of "One-Shot Adaptation of GAN in Just One CLIP" accepted to Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

Environment

Pytorch 1.7.1, Python 3.6

$ conda create -n oneshotCLIP python=3.6
$ conda install --yes -c pytorch pytorch=1.7.1 torchvision cudatoolkit=11.0
$ pip install ftfy regex tqdm
$ conda install -c anaconda git
$ conda install -c conda-forge packaging
$ pip install git+https://github.com/openai/CLIP.git

Before training, please download the pre-trained models on large datasets:

LINK: FFHQ

Training

To train the model, run

python train_oneshot.py --exp exp1 --data_path $DATA_PATH$ --ckpt $SRC_MODEL_PATH$

$DATA_PATH$ is a directory for single-shot target image

$SRC_MODEL_PATH$ is a path for source domain pre-trained model.

Default: ./pretrained_model/stylegan2-ffhq-config-f.pt

--exp is for checkpoint directory name

For human face dataset training, download portrait dataset in LINK

Testing

To test the model with adapted generator,

python test_oneshot.py --exp exp1 --ckpt $TARGET_MODEL_PATH$ --ckpt_source $SOURCE_MODEL_PATH$

$TARGET_MODEL_PATH$ is path for adapted target domain model.

$SOURCE_MODEL_PATH$ is path for source domain model. Default: ./pretrained_model/stylegan2-ffhq-config-f.pt

For testing, we provide several adapted models

LINK

Testing for real images

For testing on real images, we provide demo on Google Colab .