This is an official repository of Generating 3D Faces using Convolutional Mesh Autoencoders

[Project Page][Arxiv]

UPDATE : Thank you for using and supporting this repository over the last two years. This will no longer be maintained. Alternatively, please use:

• sw-gong/coma, thanks to Shunwang Gong.
• pixelite1201/pytorch_coma, thanks to Priyanka Patel.

This code is tested on Tensorflow 1.3. Requirements (including tensorflow) can be installed using:

pip install -r requirements.txt

Install mesh processing libraries from MPI-IS/mesh.

Download the data from the Project Page.

Preprocess the data

python processData.py --data <PATH_OF_RAW_DATA> --save_path <PATH_TO_SAVE_PROCESSED DATA>

Data pre-processing creates numpy files for the interpolation experiment and extrapolation experiment (Section X of the paper). This creates 13 different train and test files. sliced_[train|test] is for the interpolation experiment. <EXPRESSION>_[train|test] are for cross validation cross 12 different expression sequences.

To train, specify a name, and choose a particular train test split. For example,

python main.py --data data/sliced --name sliced

To test, specify a name, and data. For example,

python main.py --data data/sliced --name sliced --mode test

Run the following script. The models are slightly better (~1% on average) than ones reported in the paper.

sh generateErrors.sh

To sample faces from the latent space, specify a model and data. For example,

python main.py --data data/sliced --name sliced --mode latent

A face template pops up. You can then use the keys qwertyui to sample faces by moving forward in each of the 8 latent dimensions. Use asdfghjk to move backward in the latent space.

For more flexible usage, refer to lib/visualize_latent_space.py.

We thank Raffi Enficiaud and Ahmed Osman for pushing the release of psbody.mesh, an essential dependency for this project.

The code contained in this repository is under MIT License and is free for commercial and non-commercial purposes. The dependencies, in particular, MPI-IS/mesh and our data have their own license terms which can be found on their respective webpages. The dependencies and data are NOT covered by MIT License associated with this repository.

CAPE (CVPR 2020): Based on CoMA, we build a conditional Mesh-VAE-GAN to learn the clothing deformation from the SMPL body model, making a generative, animatable model of people in clothing. A large-scale mesh dataset of clothed humans in motion is also included!

Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J. Black. "Generating 3D faces using Convolutional Mesh Autoencoders." European Conference on Computer Vision (ECCV) 2018.