Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation

We present a generic image-to-image translation framework, Pixel2Style2Pixel (pSp). Our pSp framework is based on a novel encoder network that directly generates a series of style vectors which are fed into a pretrained StyleGAN generator, forming the extended W+ latent space. We first show that our encoder can directly embed real images into W+, with no additional optimization. We further introduce a dedicated identity loss which is shown to achieve improved performance in the reconstruction of an input image. We demonstrate pSp to be a simple architecture that, by leveraging a well-trained, fixed generator network, can be easily applied on a wide-range of image-to-image translation tasks. Solving these tasks through the style representation results in a global approach that does not rely on a local pixel-to-pixel correspondence and further supports multi-modal synthesis via the resampling of styles. Notably, we demonstrate that pSp can be trained to align a face image to a frontal pose without any labeled data, generate multi-modal results for ambiguous tasks such as conditional face generation from segmentation maps, and construct high-resolution images from corresponding low-resolution images.

Description

Official Implementation of our pSp paper for both training and evaluation. The pSp method extends the StyleGAN model to allow solving different image-to-image translation problems using its encoder.

Recent Updates

2020.10.04: Initial code release
2020.10.06: Add pSp toonify model (Thanks to the great work from Doron Adler and Justin Pinkney)!

Applications

StyleGAN Encoding

Here, we use pSp to find the latent code of real images in the latent domain of a pretrained StyleGAN generator.

Face Frontalization

In this application we want to generate a front-facing face from a given input image.

Conditional Image Synthesis

Here we wish to generate photo-realistic face images from ambiguous sketch images or segmentation maps. Using style-mixing, we inherently support multi-modal synthesis for a single input.

Super Resolution

Given a low-resolution input image, we generate a corresponding high-resolution image. As this too is an ambiguous task, we can use style-mixing to produce several plausible results.

Getting Started

Prerequisites

• Linux or macOS
• NVIDIA GPU + CUDA CuDNN (CPU may be possible with some modifications, but is not inherently supported)
• Python 2 or 3

Installation

• Clone this repo:

git clone https://github.com/eladrich/pixel2style2pixel.git
cd pixel2style2pixel

• Dependencies:
  We recommend running this repository using Anaconda. All dependencies for defining the environment are provided in environment/psp_env.yaml.

Inference Notebook

To help visualize the pSp framework on multiple tasks and to help you get started, we provide a Jupyter notebook found in notebooks/inference_playground.ipynb that allows one to visualize the various applications of pSp.
The notebook will download the necessary pretrained models and run inference on the images found in notebooks/images.
For the tasks of conditional image synthesis and super resolution, the notebook also demonstrates pSp's ability to perform multi-modal synthesis using style-mixing.

Pretrained Models

Please download the pre-trained models from the following links. Each pSp model contains the entire pSp architecture, including the encoder and decoder weights.

Path	Description
StyleGAN Inversion	pSp trained with the FFHQ dataset for StyleGAN inversion.
Face Frontalization	pSp trained with the FFHQ dataset for face frontalization.
Sketch to Image	pSp trained with the CelebA-HQ dataset for image synthesis from sketches.
Segmentation to Image	pSp trained with the CelebAMask-HQ dataset for image synthesis from segmentation maps.
Super Resolution	pSp trained with the CelebA-HQ dataset for super resolution (up to x32 down-sampling).
Toonify	pSp trained with the FFHQ dataset for toonification using StyleGAN generator from Doron Adler and Justin Pinkney.

If you wish to use one of the pretrained models for training or inference, you may do so using the flag --checkpoint_path.

In addition, we provide various auxiliary models needed for training your own pSp model from scratch as well as pretrained models needed for computing the ID metrics reported in the paper.

Path	Description
FFHQ StyleGAN	StyleGAN model pretrained on FFHQ taken from rosinality with 1024x1024 output resolution.
IR-SE50 Model	Pretrained IR-SE50 model taken from TreB1eN for use in our ID loss during pSp training.
CurricularFace Backbone	Pretrained CurricularFace model taken from HuangYG123 for use in ID similarity metric computation.
MTCNN	Weights for MTCNN model taken from TreB1eN for use in ID similarity metric computation. (Unpack the tar.gz to extract the 3 model weights.)

By default, we assume that all auxiliary models are downloaded and saved to the directory pretrained_models. However, you may use your own paths by changing the necessary values in configs/path_configs.py.

Training

Preparing your Data

• Currently, we provide support for numerous datasets and experiments (encoding, frontalization, etc.).
  • Refer to configs/paths_config.py to define the necessary data paths and model paths for training and evaluation.
  • Refer to configs/transforms_config.py for the transforms defined for each dataset/experiment.
  • Finally, refer to configs/data_configs.py for the source/target data paths for the train and test sets as well as the transforms.
• If you wish to experiment with your own dataset, you can simply make the necessary adjustments in
  1. data_configs.py to define your data paths.
  2. transforms_configs.py to define your own data transforms.

As an example, assume we wish to run encoding using ffhq (dataset_type=ffhq_encode). We first go to configs/paths_config.py and define:

dataset_paths = {
    'ffhq': '/path/to/ffhq/images256x256'
    'celeba_test': '/path/to/CelebAMask-HQ/test_img',
}

The transforms for the experiment are defined in the class EncodeTransforms in configs/transforms_config.py.
Finally, in configs/data_configs.py, we define:

DATASETS = {
   'ffhq_encode': {
        'transforms': transforms_config.EncodeTransforms,
        'train_source_root': dataset_paths['ffhq'],
        'train_target_root': dataset_paths['ffhq'],
        'test_source_root': dataset_paths['celeba_test'],
        'test_target_root': dataset_paths['celeba_test'],
    },
}

When defining our datasets, we will take the values in the above dictionary.

Training pSp

The main training script can be found in scripts/train.py.
Intermediate training results are saved to opts.exp_dir. This includes checkpoints, train outputs, and test outputs.
Additionally, if you have tensorboard installed, you can visualize tensorboard logs in opts.exp_dir/logs.

Training the pSp Encoder

python scripts/train.py \
--dataset_type=ffhq_encode \
--exp_dir=/path/to/experiment \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=0.1

Frontalization

python scripts/train.py \
--dataset_type=ffhq_frontalize \
--exp_dir=/path/to/experiment \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.08 \
--l2_lambda=0.001 \
--lpips_lambda_crop=0.8 \
--l2_lambda_crop=0.01 \
--id_lambda=1 \
--w_norm_lambda=0.005

Sketch to Face

python scripts/train.py \
--dataset_type=celebs_sketch_to_face \
--exp_dir=/path/to/experiment \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=0 \
--w_norm_lambda=0.005 \
--label_nc=1 \
--input_nc=1

Segmentation Map to Face

python scripts/train.py \
--dataset_type=celebs_seg_to_face \
--exp_dir=/path/to/experiment \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=0 \
--w_norm_lambda=0.005 \
--label_nc=19 \
--input_nc=19

Notice with conditional image synthesis no identity loss is utilized (i.e. --id_lambda=0)

Super Resolution

python scripts/train.py \
--dataset_type=celebs_super_resolution \
--exp_dir=/path/to/experiment \
--workers=8 \
--batch_size=8 \
--test_batch_size=8 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=0.1 \
--w_norm_lambda=0.005 \
--resize_factors=1,2,4,8,16,32

Additional Notes

• See options/train_options.py for all training-specific flags.
• See options/test_options.py for all test-specific flags.
• If you wish to resume from a specific checkpoint (e.g. a pretrained pSp model), you may do so using --checkpoint_path.
• If you wish to generate images from segmentation maps, please specify --label_nc=N and --input_nc=N where N is the number of semantic categories.
• Similarly, for generating images from sketches, please specify --label_nc=1 and --input_nc=1.
• Specifying --label_nc=0 (the default value), will directly use the RGB colors as input.

Testing

Inference

Having trained your model, you can use scripts/inference.py to apply the model on a set of images.
For example,

python scripts/inference.py \
--exp_dir=/path/to/experiment \
--checkpoint_path=experiment/checkpoints/best_model.pt \
--data_path=/path/to/test_data \
--test_batch_size=4 \
--test_workers=4 \
--couple_outputs

Additional notes to consider:

• During inference, the options used during training are loaded from the saved checkpoint and are then updated using the test options passed to the inference script. For example, there is no need to pass --dataset_type or --label_nc to the inference script, as they are taken from the loaded opts.
• When running inference for segmentation-to-image or sketch-to-image, it is highly recommend to do so with a style-mixing, as is done in the paper. This can simply be done by adding --latent_mask=8,9,10,11,12,13,14,15,16,17 when calling the script.
• When running inference for super-resolution, please provide a single down-sampling value using --resize_factors.
• Adding the flag --couple_outputs will save an additional image containing the input and output images side-by-side in the sub-directory inference_coupled. Otherwise, only the output image is saved to the sub-directory inference_results.
• By default, the images will be saved at resolutiosn of 1024x1024, the original output size of StyleGAN. If you wish to save outputs resized to resolutions of 256x256, you can do so by adding the flag --resize_outputs.

Multi-Modal Synthesis with Style-Mixing

Given a trained model for conditional image synthesis or super-resolution, we can easily generate multiple outputs for a given input image. This can be done using the script scripts/style_mixing.py.
For example, running the following command will perform style-mixing for a segmentation-to-image experiment:

python scripts/style_mixing.py \
--exp_dir=/path/to/experiment \
--checkpoint_path=/path/to/experiment/checkpoints/best_model.pt \
--data_path=/path/to/test_data/ \
--test_batch_size=4 \
--test_workers=4 \
--n_images=25 \
--n_outputs_to_generate=5 \
--latent_mask=8,9,10,11,12,13,14,15,16,17

Here, we inject 5 randomly drawn vectors and perform style-mixing on the latents [8,9,10,11,12,13,14,15,16,17].

Additional notes to consider:

• To perform style-mixing on a subset of images, you may use the flag --n_images. The default value of None will perform style mixing on every image in the given data_path.
• You may also include the argument --mix_alpha=m where m is a float defining the mixing coefficient between the input latent and the randomly drawn latent.
• When performing style-mixing for super-resolution, please provide a single down-sampling value using --resize_factors.
• By default, the images will be saved at resolutiosn of 1024x1024, the original output size of StyleGAN. If you wish to save outputs resized to resolutions of 256x256, you can do so by adding the flag --resize_outputs.

Computing Metrics

Similarly, given a trained model and generated outputs, we can compute the loss metrics on a given dataset.
These scripts receive the inference output directory and ground truth directory.

• Calculating the identity loss:

python scripts/calc_id_loss_parallel.py \
--data_path=/path/to/experiment/inference_outputs \
--gt_path=/path/to/test_images \

• Calculating LPIPS loss:

python scripts/calc_losses_on_images.py \
--mode lpips
--data_path=/path/to/experiment/inference_outputs \
--gt_path=/path/to/test_images \

• Calculating L2 loss:

python scripts/calc_losses_on_images.py \
--mode l2
--data_path=/path/to/experiment/inference_outputs \
--gt_path=/path/to/test_images \

Additional Applications

To better show the flexibility of our pSp framework we present additional applications below.

As with our main applications, you may download the pretrained models here:

Path	Description
Toonify	pSp trained with the FFHQ dataset for toonification using StyleGAN generator from Doron Adler and Justin Pinkney.

Toonify

Using the toonify StyleGAN built by Doron Adler and Justin Pinkney, we take a real face image and generate a toonified version of the given image. We train the pSp encoder to directly reconstruct real face images inside the toons latent space resulting in a projection of each image to the closest toon. We do so without requiring any labeled pairs or distillation!

This is trained exactly like the StyleGAN inversion task with several changes:

• Change from FFHQ StyleGAN to toonifed StyleGAN (can be set using --stylegan_weights)
  • The toonify generator is taken from Doron Adler and Justin Pinkney and converted to Pytorch using rosinality's conversion script.
  • For convenience, the converted generator Pytorch model may be downloaded here.
• Increase id_lambda from 0.1 to 1
• Increase w_norm_lambda from 0.005 to 0.025

We obtain the best results after around 6000 iterations of training (can be set using --max_steps)

Repository structure

Path	Description
pixel2style2pixel	Repository root folder
├  configs	Folder containing configs defining model/data paths and data transforms
├  criteria	Folder containing various loss criterias for training
├  datasets	Folder with various dataset objects and augmentations
├  environment	Folder containing Anaconda environment used in our experiments
├ models	Folder containting all the models and training objects
│  ├  encoders	Folder containing our pSp encoder architecture implementation and ArcFace encoder implementation from TreB1eN
│  ├  mtcnn	MTCNN implementation from TreB1eN
│  ├  stylegan2	StyleGAN2 model from rosinality
│  └  psp.py	Implementation of our pSp framework
├  notebook	Folder with jupyter notebook containing pSp inference playground
├  options	Folder with training and test command-line options
├  scripts	Folder with running scripts for training and inference
├  training	Folder with main training logic and Ranger implementation from lessw2020
├  utils	Folder with various utility functions

TODOs

• Add multi-gpu support

Citation

If you use this code for your research, please cite our paper Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation:

@article{richardson2020encoding,
  title={Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation},
  author={Richardson, Elad and Alaluf, Yuval and Patashnik, Or and Nitzan, Yotam and Azar, Yaniv and Shapiro, Stav and Cohen-Or, Daniel},
  journal={arXiv preprint arXiv:2008.00951},
  year={2020}
}