1. Getting Started
2. Experiments
3. Acknowledgments
4. Contacts

Tan M. Dinh, Anh Tran, Rang Nguyen, Binh-Son Hua
VinAI Research, Vietnam

Abstract: Real-world image manipulation has achieved fantastic progress in recent years as a result of the exploration and utilization of GAN latent spaces. GAN inversion is the first step in this pipeline, which aims to map the real image to the latent code faithfully. Unfortunately, the majority of existing GAN inversion methods fail to meet at least one of the three requirements listed below: high reconstruction quality, editability, and fast inference. We present a novel two-phase strategy in this research that fits all requirements at the same time. In the first phase, we train an encoder to map the input image to StyleGAN2 W space, which was proven to have excellent editability but lower reconstruction quality. In the second phase, we supplement the reconstruction ability in the initial phase by leveraging a series of hypernetworks to recover the missing information during inversion. These two steps complement each other to yield high reconstruction quality thanks to the hypernetwork branch and excellent editability due to the inversion done in the W space. Our method is entirely encoder-based, resulting in extremely fast inference. Extensive experiments on two challenging datasets demonstrate the superiority of our method.

Our method outperforms significantly other encoder-based methods (pSp, e4e, ReStyle) while having the inferece time as same as them. In comparison with optimization-based approaches (SG2-W+, PTI), our work is on par of quality with SG2-W+, and sightly lower than PTI. However, it is worth noting that, our method run very faster allowing for interactive applications (3000 and 1100 times shorter than SG2-W+ and PTI).

Details of the model architecture and experimental results can be found in our following paper.

@inproceedings{dinh2021hyperinverter,
    title={HyperInverter: Improving StyleGAN Inversion via Hypernetwork},
    author={Tan M. Dinh and Anh Tuan Tran and Rang Nguyen and Binh-Son Hua},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2022}
}

Please CITE our paper whenever our model implementation is used to help produce published results or incorporated into other software.

The codebase is tested on

• Ubuntu
• CUDA 10.0, CuDNN 7

• Clone this repo:

git clone https://github.com/VinAIResearch/HyperInverter.git
cd HyperInverter

• Install dependencies:

conda create -p ./envs python=3.7.3
conda activate ./envs
pip install -r requirements.txt

• Human Faces: We use 70,000 images from FFHQ dataset to train, and 2,824 images from CelebA-HQ dataset to test. The images have 1024 x 1024 resolution and are cropped and aligned to the center. Refering FFHQ for more details about the pre-processing step.

• Churches: We use all 126,227 images and 300 images from the official train and test sets of LSUN Church to train and evaluate our model. The images are resized to 256 x 256 resolution.

Please download the corresponding datasets and unzip to the data folder. Then, go to configs/paths_config.py and modify this file to link data correctly:

dataset_paths = {
    "ffhq": "/path/to/ffhq/train_img",
    "celeba_test": "/path/to/CelebA-HQ/test_img",
    "church_train": "/path/to/lsun-church/train_img",
    "church_test": "/path/to/lsun-church/test_img",
}

If you need to try with your own dataset, you can make the necessary modifications in: (i) data_configs.py to define your data paths; (ii) transforms_configs.py to define your data transformations.

Run the below command to automatically download the auxiliary pre-trained models needed in the experiments.

python scripts/download_auxiliary_pretrained_models.py

Basicly, we set these paths correctly to training or inference processes. If you want to change these paths, please visit file configs/paths_config.py and modify the corresponding record in the model_paths dict.

See Model Zoo for our official pre-trained models. Please download our pretrained models (both W Encoder and HyperInverter) and put them to the pretrained_models folder.

Please follow pSp or e4e to train the W encoder to encode images to latent codes in W space (512).

To save time, we release our pre-trained W encoders for churches and human faces in Model Zoo.

We provide the default training scripts as below. If you have time, please tune the hyper-parameters further to get the best results. Note that the argument --hidden_dim has a large effect on the quality of the model, increasing them cause a bigger model and tend to have better quality. In the paper, we use hidden_dim=256. However, we recommend using hidden_dim=128 to balance the model size and model performance. In addition, it is worth noting that, our code is also easy to modify for predicting different layer weights of StyleGAN2. Refering to weight_shapes for reference and modify them if you need to try with other layers. Then, you modify the get_target_shapes function in hyper_inverter to add your option.

• Human Faces

EXPERIMENT_DIR=""
W_ENCODER_PATH=""
GPU_ID=0

CUDA_VISIBLE_DEVICES="$GPU_ID" \
python scripts/train.py \
--dataset_type=ffhq_encode \
--encoder_type=LayerWiseEncoder \
--w_encoder_path="$W_ENCODER_PATH" \
--output_size=1024 \
--exp_dir="$EXPERIMENT_DIR" \
--batch_size=8 \
--batch_size_used_with_adv_loss=4 \
--workers=4 \
--val_interval=1000 \
--save_interval=5000 \
--encoder_optim_name=adam \
--discriminator_optim_name=adam \
--encoder_learning_rate=1e-4 \
--discriminator_learning_rate=1e-4 \
--hyper_lpips_lambda=0.8 \
--hyper_l2_lambda=1.0 \
--hyper_id_lambda=0.1 \
--hyper_adv_lambda=0.005 \
--hyper_d_reg_every=16 \
--hyper_d_r1_gamma=10.0 \
--step_to_add_adversarial_loss=200000 \
--target_shape_name=conv_without_bias  \
--max_steps=500000 \
--hidden_dim=128 \
--num_cold_steps=20000 \
--save_checkpoint_for_resuming_training \
--use_wandb

• Churches

EXPERIMENT_DIR=""
W_ENCODER_PATH=""
GPU_ID=0

CUDA_VISIBLE_DEVICES="$GPU_ID" \
python scripts/train.py \
--dataset_type=church_encode \
--encoder_type=ResNetLayerWiseEncoder \
--w_encoder_path="$W_ENCODER_PATH" \
--output_size=256 \
--exp_dir="$EXPERIMENT_DIR" \
--batch_size=8 \
--batch_size_used_with_adv_loss=4 \
--workers=4 \
--val_interval=1000 \
--save_interval=5000 \
--encoder_optim_name=adam \
--discriminator_optim_name=adam \
--encoder_learning_rate=1e-4 \
--discriminator_learning_rate=1e-4 \
--hyper_lpips_lambda=0.8 \
--hyper_l2_lambda=1.0 \
--hyper_id_lambda=0.5 \
--hyper_adv_lambda=0.15 \
--hyper_d_reg_every=16 \
--hyper_d_r1_gamma=100.0 \
--step_to_add_adversarial_loss=100000 \
--target_shape_name=conv_without_bias \
--max_steps=500000 \
--hidden_dim=128 \
--num_cold_steps=10000 \
--save_checkpoint_for_resuming_training \
--use_wandb

• Placing the input images in a folder.

• Pre-processing (if needed). For human facial domain, if the input images have not been cropped and aligned, please run the below script to preprocess data.

RAW_IMAGE_DIR=""
PROCESSED_IMAGE_DIR=""

python scripts/align_all_parallel.py \
--raw_dir "$RAW_IMAGE_DIR" \
--saved_dir "$PROCESSED_IMAGE_DIR" \
--num_threads 8 

The descriptions of the arguments are shown below.

• Run inference

Set the argument correctly before running the below script.

INPUT_DATA_DIR=""
RESULT_DIR=""
MODEL_PATH=""
GPU_ID=0

CUDA_VISIBLE_DEVICES="$GPU_ID" \
python scripts/inference.py \
--exp_dir="$RESULT_DIR" \
--checkpoint_path="$MODEL_PATH" \
--data_path="$INPUT_DATA_DIR" \
--batch_size=4 \
--workers=4

The description of the arguments are shown below.

Finally, the reconstructed images can be found in RESULT_DIR/inference_results folder.

We have prepared the scripts for quantitatively reconstruction evaluation on human faces and churches models. Please set the arguments to compatible with your model in these files and run below commands to conduct the evaluation process.

• Human Faces

sh sample_scripts/human_faces_reconstruction_quantitative_evaluation.sh

• Churches

sh sample_scripts/church_reconstruction_quantitative_evaluation.sh

For the following experiments, please visit file configs/paths_config.py and update model_paths dict with the paths to the pre-trained models of HyperInverter and other inversion methods. For other inversion methods, please visit their Github repositories to download their pre-trained weights.

The sample script for qualitatively reconstruction comparison is:

DOMAIN=""
METHODS=""
INPUT_DATA_DIR=""
SAVED_RESULTS_DIR_NAME=""
MAX_NUM_IMAGES=100
SAVED_SIZE=1024
GPU_ID=0


CUDA_VISIBLE_DEVICES="$GPU_ID" \
python evaluation/reconstruction_comparison.py \
--methods="$METHODS" \
--domain="$DOMAIN" \
--input_data_dir="$INPUT_DATA_DIR"  \
--input_data_id="$SAVED_RESULTS_DIR_NAME" \
--output_dir=outputs \
--saved_embedding_dir=embeddings \
--max_num_images="$MAX_NUM_IMAGES" \
--resize="$SAVED_SIZE"

The descriptions of the arguments are shown below.

The results can be found in the outputs/SAVED_RESULTS_DIR_NAME folder.

The sample script for editing comparison is:

DOMAIN=""
METHODS=""
DIRECTION=""
INPUT_DATA_DIR=""
SAVED_RESULTS_DIR_NAME=""
MAX_NUM_IMAGES=10
SAVED_SIZE=1024
MIN_MAG=-30  
MAX_MAG=30
STEP=5 
GPU_ID=0

CUDA_VISIBLE_DEVICES="$GPU_ID" \
python evaluation/editing_inference.py \
--methods="$METHODS" \
--domain="$DOMAIN" \
--input_data_dir="$INPUT_DATA_DIR" \
--input_data_id="$SAVED_RESULTS_DIR_NAME" \
--output_dir=outputs \
--saved_embedding_dir=embeddings \
--direction="$DIRECTION" \
--min_factor="$MIN_MAG" \
--max_factor="$MAX_MAG" \
--step="$STEP" \
--max_num_images="$MAX_NUM_IMAGES" \
--resize="$SAVED_SIZE" \
--save_edited_images \
--gif_speed=4 

The results can be found in the outputs/SAVED_RESULTS_DIR_NAME folder. Please try with different values of MIN_MAG and MAX_MAG to get the best result.

The descriptions of the arguments are shown below.

The supported editing directions are listed below.

• Human Faces

• Churches

The sample script for the interpolation of two real images is:

DOMAIN=""
METHODS=""
PATH_TO_INPUT_IMAGE_1=""
PATH_TO_INPUT_IMAGE_2=""
SAVED_RESULTS_DIR=""
SAVED_RESULTS_FILE_NAME=""
SAVED_SIZE=1024
NUM_STEPS=100
GPU_ID=0

CUDA_VISIBLE_DEVICES="$GPU_ID" \
python evaluation/real_image_interpolation.py \
--domain="$DOMAIN" \
--method="$METHODS" \
--left_image_path="$PATH_TO_INPUT_IMAGE_1" \
--right_image_path="$PATH_TO_INPUT_IMAGE_2" \
--steps="$NUM_STEPS" \
--saved_image_size="$SAVED_SIZE" \
--saved_dir="$SAVED_RESULTS_DIR" \
--saved_file_name="$SAVED_RESULTS_FILE_NAME" \
--save_interpolated_images \
--gif_speed=2

The descriptions of the arguments are shown below.

Our source code is developed based on the codebase of a great series of StyleGAN inversion researches from the Tel Aviv University group, which are: pSp, e4e, ReStyle and PTI.

For auxiliary pre-trained models, we specifically thank to TreB1eN, MoCov2, CurricularFace and MTCNN. For editing directions, thanks to the authors of GANSpace, InterFaceGAN and StyleCLIP.

We leverage the PyTorch implementation of StyleGAN2-ADA for the StyleGAN model. All pre-trained StyleGAN models are from the official release of StyleGAN2. We convert the original weights exported by TensorFlow code to compatible with the PyTorch version of StyleGAN2-ADA by using the author's official script.

Overall, thank you so much to the authors for their great works and efforts to release source code and pre-trained weights.

If you have any questions, please drop an email to [email protected] or open an issue in this repository.