Semantic Deep Face Models

From Semantic Deep Face Models by Prashanth Chandran et al. (2020)

This repository implements the paper Semantic Deep Face Models by Prashanth Chandran et al. The albedo generation is omitted. The system uses the Facewarehouse dataset to train the VAE. The goal of the system is to build an autoencoder that generates new faces. It disentangles identity and expression by design and provides two latent spaces (Z_id and Z_exp) that can be tweaked separately to modify expression or identity. The system uses fully connected layers in the encoder and decoder with residual connections that span across the entire achitecture. One improvement would be to represent the meshes with graph convolutions instead.

Installation

Create a new conda environment (optional):

conda env create -n <env_name> python=3.10
conda activate <env_name>

Install dependencies:

pip install -r requirements.txt

If you are on Windows, you can install CUDA for faster training:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Run

Train model: python train.py
Random faces: python sample.py
Interpolation: python interpolate.py

Requirements

• Python 3.10.x
• torch
• numpy
• trimesh
• pyglet
• scipy
• matplotlib
• opengl
• pyqt6