JackZhouSz / EMOPortraits

Official implementation of EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars

EMOPortraits

Official implementation of EMOPortraits: Emotion-enhanced Multimodal One-shot Head Avatars.

Overview

EMOPortraits introduces a novel approach for generating realistic and expressive one-shot head avatars driven by multimodal inputs, including extreme and asymmetric emotions.

For more details, please refer to:

• Project Page
• Research Paper
• FEED Dataset Repository
• Main author's Page

---

Table of Contents

• Overview
• Installation
  • Environment Setup
  • Installing Additional Repositories
  • Downloading Pre-trained Models and Dependencies
• Usage
• Notes
  • Note 1: Repository Maintenance
  • Note 2: Importance of High-Quality Data
  • Note 3: Utilizing the FEED Dataset
  • Note 4: Model Weights
• Acknowledgements
• License

---

Installation

1. Environment Setup

You can set up the environment using the provided conda-pack archive:

Option 1: Use conda-pack Archive

1. Download the Environment Archive

   Download sav.tar.gz from the Google Drive or Yandex Disk.

2. Unpack the Environment

   # Create a directory for the environment
   mkdir -p sav_env
   
   # Unpack the tar.gz archive into the directory
   tar -xzf sav.tar.gz -C sav_env

3. Using Python Without Activating

   # Run Python directly from the unpacked environment
   ./sav_env/bin/python

4. Activating the Environment

   # Activate the environment
   source sav_env/bin/activate

   Once activated, you can run Python as usual:

   (sav_env) $ python

5. Cleanup Prefixes

   After activating the environment, you may need to run the following command to fix any issues with environment paths:

   (sav_env) $ conda-unpack

   This command can also be run without activating the environment, as long as Python is installed on the machine.

Option 2: Using environment.yml

Note: This option may not work as it has not been thoroughly tested.

2. Installing Additional Repositories

Due to limitations with conda-pack, the following repositories need to be installed manually:

1. Face Detection: Install from GitHub

   git clone https://github.com/hhj1897/face_detection.git
   cd face_detection
   pip install -e .

2. ROI Tanh Warping: Install from GitHub

   git clone https://github.com/ibug-group/roi_tanh_warping.git
   cd roi_tanh_warping
   pip install -e .

3. Face Parsing: Install from GitHub

   git clone https://github.com/hhj1897/face_parsing.git
   cd face_parsing
   pip install -e .

3. Downloading Pre-trained Models and Dependencies

1. Download Required Files

   Please download the following files from Google Drive or Yandex Disk:

   • logs.zip (contains main model weights - not yet available)
   • logs_s2.zip (contains stage 2 model weights)
   • repos.zip (contains dependencies repos and it's weights)

2. Extract Files

   Extract all the downloaded zip files into the root directory of the project:

   unzip logs.zip -d ./
   unzip logs_s2.zip -d ./
   unzip repos.zip -d ./

3. Download and Extract Loss Models

   Navigate to the losses directory and download the following files:

   cd losses

   • loss_model_weights.zip
   • gaze_models.zip

   Extract them within the same losses directory:

   unzip loss_model_weights.zip -d ./
   unzip gaze_models.zip -d ./

---

Usage

Instructions on how to run the code, train models, and perform inference will be added here.

---

Notes

Note 1: Repository Maintenance

This repository is primarily intended for demonstration purposes, allowing enthusiasts to explore the network architecture and training procedures in detail. The primary author is not currently affiliated with academia and may not have the capacity to actively maintain this repository. Community contributions and support are highly encouraged.

Note 2: Importance of High-Quality Data

A significant factor contributing to the success and quality of the results is the dataset used for training. The original model was trained on a high-quality (HQ) version of the VoxCeleb2 dataset, which is no longer publicly available. However, there are now newer datasets of higher quality and larger scale. Utilizing these can potentially yield even better results, as seen in recent methods that build upon ideas presented in the MegaPortraits paper.

Note 3: Utilizing the FEED Dataset

Our FEED dataset (link), introduced in our paper, was instrumental in incorporating asymmetric and extreme emotions into the latent emotion space. We encourage the community to actively use and expand upon this dataset. Given that the final version is slightly smaller (due to some participants withdrawing consent), supplementing it with other datasets containing extreme emotions (e.g., NeRSemble) can enhance model performance, especially when attempting to replicate or improve upon the techniques presented in EMOPortraits.

Note 4: Model Weights

We plan to provide version of the pre-trained model weights soon (will be located in logs.zip):

1. Retrain_with_17_V1_New_rand_MM_SEC_4_drop_02_stm_10_CV_05_1_1

This model will be retrained using the same parameters as described in our paper but with 17 IDs in the FEED dataset instead of the original 23. Since the FEED dataset samples were used 25% of the time during training, this change might slightly affect performance in intensive tests.

---

Acknowledgements

We extend our gratitude to all contributors and participants who made this project possible. Special thanks to the developers of the datasets and tools that were instrumental in our research.

---

License

This project is licensed under the Creative Commons BY-NC-SA 4.0 license. You are free to use, modify, and distribute this work non-commercially, as long as appropriate credit is given and any derivative works are licensed under identical terms.