Frequency Domain Image Translation: More Photo-realistic, Better Identity-preserving

This is the source code for our paper Frequency Domain Image Translation: More Photo-realistic, Better Identity-preserving by Mu Cai, Hong Zhang, Huijuan Huang, Qichuan Geng, Yixuan Li and Gao Huang. Code is modified from Swapping Autoencoder, StarGAN v2, Image2StyleGAN.

This is a frequency-based image translation framework that is effective for identity preserving and image realism. Our key idea is to decompose the image into low-frequency and high-frequency components, where the high-frequency feature captures object structure akin to the identity. Our training objective facilitates the preservation of frequency information in both pixel space and Fourier spectral space.

1. Swapping Autoencoder

Dataset Preparation

You can download the following datasets:

• Flicker Mountain (training set, validation set)
• Flicker Waterfall (training set, validation set)
• CelebA-HQ
• LSUN Church
• LSUN Bedroom

Then place the training data and validation data in ./swapping-autoencoder/dataset/.

Train the model

You can train the model using either lmdb or folder format. For training the FDIT assisted Swapping Autoencoder, please run:

cd swapping-autoencoder 
bash train.sh

Change the location of the dataset according to your own setting.

Evaluate the model

Generate image hybrids

Place the source images and reference images under the folder ./sample_pair/source and ./sample_pair/ref respectively. The two image pairs should have the exact same index, such as 0.png, 1.png, ...

To generate the image hybrids according to the source and reference images, please run:

bash eval_pairs.sh

Evaluate the image quality

To evaluate the image quality using Fréchet Inception Distance (FID), please run

bash eval.sh

The pretrained model is provided here.

2. Image2StyleGAN

Prepare the dataset

You can place your own images or our official dataset under the folder ./Image2StlyleGAN/source_image. If using our dataset, then unzip it into that folder.

cd Image2StlyleGAN
unzip source_image.zip 

Get the weight files

To get the pretrained weights in StyleGAN, please run:

cd Image2StlyleGAN/weight_files/pytorch

Download the weight file through this link:

Run GAN-inversion model:

Single image inversion

Run the following command by specifying the name of the image image_name:

python encode_image_freq.py --src_im  image_name

Group images inversion

Please run

python encode_image_freq_batch.py 

Quantitative Evaluation

To get the image reconstruction metrics such as MSE, MAE, PSNR, please run:

python eval.py         

3. StarGAN v2

Prepare the dataset

Please download the CelebA-HQ-Smile dataset into ./StarGANv2/data

Train the model

To train the model in Tesla V100, please run:

cd StarGANv2
bash train.sh

Evaluation

To get the image translation samples and image quality measures like FID, please run:

bash eval.sh

Pretrained Model

The pretrained model can be found here.

Image Translation Results

FDIT achieves state-of-the-art performance in several image translation and even GAN-inversion models.

Citation

If you use our codebase or datasets, please cite our work:

@article{cai2021frequency,
title={Frequency Domain Image Translation: More Photo-realistic, Better Identity-preserving},
author={Cai, Mu and Zhang, Hong and Huang, Huijuan and Geng, Qichuan and Li, Yixuan and Huang, Gao},
journal={In Proceedings of International Conference on Computer Vision (ICCV)},
year={2021}
}