BDInvert: GAN Inversion for Out-of-Range Images with Geometric Transformations

Prerequisites

pip install -r requirements.txt

cd BDInvert

python train_basecode_encoder.py

--model_name : You can change backbone StyleGAN model, which located at model zoo.
--basecode_spatial_size : You can change spatial resolution of basecode.

python pca_p_space.py

Download and unzip under BDInvert/pretrained_models/.

Encoder Pretrained Models	Basc Code Spatial Size
StyleGAN2 pretrained on FFHQ 1024, 16x16	16x16

Default test setting use StyleGAN2 pretrained on FFHQ1024 and use basecode spatial size as 16x16.

cd BDInvert

python make_list.py --image_folder ./test_img

python invert.py --encoder_pt_path {encoder_pt_path}

--image_list : Inversion target image list generated from above step 2. Default is ./test_img/test.list
--weight_pnorm_term : As recently well known, there is a trade-off between editing quality and reconstruction quality. This argument controls this trade-off.

python edit.py {inversion directory}

--edit_direction : You can change edit direction, which located at BDInvert/editings/interfacegan_directions

We changed the detail code regularization method from hard clipping in P-norm+ space to L2 norm regularization, following the update of the original paper.
Due to this change, new hyperparameter, weight_pnorm_term, has been added.

This software is being made available under the terms in the LICENSE file.

Any exemptions to these terms requires a license from the Pohang University of Science and Technology.

NOTE