Cycle Encoding of a StyleGAN Encoder for Improved Reconstruction and Editability

November 3, 2022 · View on GitHub

Description

Official Implementation of "Cycle Encoding of a StyleGAN Encoder for Improved Reconstruction and Editability" paper.

Prerequisites and Installation

Please refer to the Prerequisites and Installation in PTI.

Pretrained Models

Please download the pre-trained models from the following links. We assume that all auxiliary models are downloaded and saved to the directory pretrained_models.

Link
FFHQ CycleEncoding Inversion

Please also download the auxiliary models from e4e.

Inversion

Method 1: Cycle Encoding + Pivotal Tuning

As described in the paper, this method is faster than the second method. Please perform the following steps:

cd pivotal_tuning

Edit "configs/hyperparameters.py" and set "use_saved_w_pivots = False" and "first_inv_type = 'cycle'".

sh run_pivotal_tuning.sh

This method achieves better reconstruction quality than the first method. Please perform the following steps:

cd refinement
sh run_regularized_refinement.sh

Copy or link the directory saved_w to the directory pivotal_tuning.

cd pivotal_tuning

Edit "configs/hyperparameters.py" and set "use_saved_w_pivots = True".

sh run_pivotal_tuning.sh

Training

Train W -> W+

cd cycle_encoding/w_to_wplus
sh run_w_to_wplus.sh

Train W+ -> W

cd cycle_encoding/wplus_to_w
sh run_wplus_to_w.sh

Quantitative Evaluation

We used the scripts calc_id_loss_parallel.py and calc_losses_on_images.py from pSp for quantitative evaluation.

Acknowledgments

The code borrows heavily from PTI. Some code borrows from e4e and pSp.

Citation

If you use this code for your research, please cite our paper:

@inproceedings{Xudong2022CycleEncoding,
  title={Cycle Encoding of a StyleGAN Encoder for Improved Reconstruction and Editability},
  author={Xudong Mao and Liujuan Cao and Aurele T. Gnanha and Zhenguo Yang and Qing Li and Rongrong Ji},
  booktitle={Proceedings of ACM International Conference on Multimedia},
  year={2022}
}