PTI: Pivotal Tuning for Latent-based editing of Real Images (ACM TOG 2022)

<!-- > Recently, a surge of advanced facial editing techniques have been proposed that leverage the generative power of a pre-trained StyleGAN. To successfully edit an image this way, one must first project (or invert) the image into the pre-trained generator’s domain. As it turns out, however, StyleGAN’s latent space induces an inherent tradeoff between distortion and editability, i.e. between maintaining the original appearance and convincingly altering some of its attributes. Practically, this means it is still challenging to apply ID-preserving facial latent-space editing to faces which are out of the generator’s domain. In this paper, we present an approach to bridge this gap. Our technique slightly alters the generator, so that an out-of-domain image is faithfully mapped into an in-domain latent code. The key idea is pivotal tuning — a brief training process that preserves the editing quality of an in-domain latent region, while changing its portrayed identity and appearance. In Pivotal Tuning Inversion (PTI), an initial inverted latent code serves as a pivot, around which the generator is fined-tuned. At the same time, a regularization term keeps nearby identities intact, to locally contain the effect. This surgical training process ends up altering appearance features that represent mostly identity, without affecting editing capabilities. To supplement this, we further show that pivotal tuning can also adjust the generator to accommodate a multitude of faces, while introducing negligible distortion on the rest of the domain. We validate our technique through inversion and editing metrics, and show preferable scores to state-of-the-art methods. We further qualitatively demonstrate our technique by applying advanced edits (such as pose, age, or expression) to numerous images of well-known and recognizable identities. Finally, we demonstrate resilience to harder cases, including heavy make-up, elaborate hairstyles and/or headwear, which otherwise could not have been successfully inverted and edited by state-of-the-art methods. -->

<a href="https://arxiv.org/abs/2106.05744"><img src="https://img.shields.io/badge/arXiv-2008.00951-b31b1b.svg"></a> <a href="https://opensource.org/licenses/MIT"><img src="https://img.shields.io/badge/License-MIT-yellow.svg"></a>
Inference Notebook: <a href="https://colab.research.google.com/github/danielroich/PTI/blob/main/notebooks/inference_playground.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" height=20></a>

<p align="center"> <img src="docs/teaser.jpg"/> <br> Pivotal Tuning Inversion (PTI) enables employing off-the-shelf latent based semantic editing techniques on real images using StyleGAN. PTI excels in identity preserving edits, portrayed through recognizable figures — Serena Williams and Robert Downey Jr. (top), and in handling faces which are clearly out-of-domain, e.g., due to heavy makeup (bottom). </br> </p>

Description

Official Implementation of our PTI paper + code for evaluation metrics. PTI introduces an optimization mechanizem for solving the StyleGAN inversion task. Providing near-perfect reconstruction results while maintaining the high editing abilitis of the native StyleGAN latent space W. For more details, see <a href="https://arxiv.org/abs/2106.05744"><img src="https://img.shields.io/badge/arXiv-2008.00951-b31b1b.svg"></a>

Recent Updates

2021.07.01: Fixed files download phase in the inference notebook. Which might caused the notebook not to run smoothly.

2021.06.29: Added support for CPU. In order to run PTI on CPU please change device parameter under configs/global_config.py to "cpu" instead of "cuda".

2021.06.25 : Adding mohawk edit using StyleCLIP+PTI in inference notebook. Updating documentation in inference notebook due to Google Drive rate limit reached. Currently, Google Drive does not allow to download the pretrined models using Colab automatically. Manual intervention might be needed.

Getting Started

Prerequisites

• Linux or macOS
• NVIDIA GPU + CUDA CuDNN (Not mandatory bur recommended)
• Python 3

Installation

• Dependencies:
  1. lpips
  2. wandb
  3. pytorch
  4. torchvision
  5. matplotlib
  6. dlib
• All dependencies can be installed using pip install and the package name

Pretrained Models

Please download the pretrained models from the following links.

Auxiliary Models

We provide various auxiliary models needed for PTI inversion task.
This includes the StyleGAN generator and pre-trained models used for loss computation. | Path | Description | :--- | :---------- |FFHQ StyleGAN | StyleGAN2-ada model trained on FFHQ with 1024x1024 output resolution. |Dlib alignment | Dlib alignment used for images preproccessing. |FFHQ e4e encoder | Pretrained e4e encoder. Used for StyleCLIP editing.

Note: The StyleGAN model is used directly from the official stylegan2-ada-pytorch implementation. For StyleCLIP pretrained mappers, please see StyleCLIP's official routes

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.

Inversion

Preparing your Data

In order to invert a real image and edit it you should first align and crop it to the correct size. To do so you should perform One of the following steps:

1. Run notebooks/align_data.ipynb and change the "images_path" variable to the raw images path
2. Run utils/align_data.py and change the "images_path" variable to the raw images path

Weights And Biases

The project supports Weights And Biases framework for experiment tracking. For the inversion task it enables visualization of the losses progression and the generator intermediate results during the initial inversion and the Pivotal Tuning(PT) procedure.

The log frequency can be adjusted using the parameters defined at configs/global_config.py under the "Logs" subsection.

There is no no need to have an account. However, in order to use the features provided by Weights and Biases you first have to register on their site.

Running PTI

The main training script is scripts/run_pti.py. The script receives aligned and cropped images from paths configured in the "Input info" subscetion in configs/paths_config.py. Results are saved to directories found at "Dirs for output files" under configs/paths_config.py. This includes inversion latent codes and tuned generators. The hyperparametrs for the inversion task can be found at configs/hyperparameters.py. They are intilized to the default values used in the paper.

Editing

By default, we assume that all auxiliary edit directions are downloaded and saved to the directory editings. However, you may use your own paths by changing the necessary values in configs/path_configs.py under "Edit directions" subsection.

Example of editing code can be found at scripts/latent_editor_wrapper.py

Inference Notebooks

To help visualize the results of PTI we provide a Jupyter notebook found in notebooks/inference_playground.ipynb.
The notebook will download the pretrained models and run inference on a sample image found online or on images of your choosing. It is recommended to run this in Google Colab.

The notebook demonstrates how to:

• Invert an image using PTI
• Visualise the inversion and use the PTI output
• Edit the image after PTI using InterfaceGAN and StyleCLIP
• Compare to other inversion methods

Evaluation

Currently the repository supports qualitative evaluation for reconstruction of: PTI, SG2 (W Space), e4e, SG2Plus (W+ Space). As well as editing using InterfaceGAN and GANSpace for the same inversion methods. To run the evaluation please see evaluation/qualitative_edit_comparison.py. Examples of the evaluation scripts are:

<p align="center"> <img src="docs/model_rec.jpg"/> <br> Reconsturction comparison between different methods. The images order is: Original image, W+ inversion, e4e inversion, W inversion, PTI inversion </br> </p> <p align="center"> <img src="docs/stern_rotation.jpg"/> <br> InterfaceGAN pose edit comparison between different methods. The images order is: Original, W+, e4e, W, PTI </br> </p> <p align="center"> <img src="docs/tyron_original.jpg" width="220" height="220"/> <img src="docs/tyron_edit.jpg" width="220" height="220"/> <br> Image per edit or several edits without comparison </br> </p>

Coming Soon - Quantitative evaluation and StyleCLIP qualitative evaluation

Repository structure

| Path | Description <img width=200> | :--- | :--- | ├  configs | Folder containing configs defining Hyperparameters, paths and logging | ├  criteria | Folder containing various loss and regularization criterias for the optimization | ├  dnnlib | Folder containing internal utils for StyleGAN2-ada | ├  docs | Folder containing the latent space edit directions | ├  editings | Folder containing images displayed in the README | ├  environment | Folder containing Anaconda environment used in our experiments | ├  licenses | Folder containing licenses of the open source projects used in this repository | ├  models | Folder containing models used in different editing techniques and first phase inversion | ├  note