This repository contains the code to run the trained models of SNUG.

We present a self-supervised method to learn dynamic 3D deformations of garments worn by parametric human bodies. State-of-the-art data-driven approaches to model 3D garment deformations are trained using supervised strategies that require large datasets, usually obtained by expensive physics-based simulation methods or professional multi-camera capture setups. In contrast, we propose a new training scheme that removes the need for ground-truth samples, enabling self-supervised training of dynamic 3D garment deformations. Our key contribution is to realize that physics-based deformation models, traditionally solved in a frame-by-frame basis by implicit integrators, can be recasted as an optimization problem. We leverage such optimization-based scheme to formulate a set of physics-based loss terms that can be used to train neural networks without precomputing ground-truth data. This allows us to learn models for interactive garments, including dynamic deformations and fine wrinkles, with a two orders of magnitude speed up in training time compared to state-of-the-art supervised methods.

Requirements: python3.8, tensorflow-2.7, numpy-1.21.4, scipy-1.7.2

(Recommended) Create virtual environment:

python3 -m venv venv
source venv/bin/activate // Ubuntu
.\venv\Scripts\activate  // Windows
pip install -r requirements.txt

1. Sign in into https://smpl.is.tue.mpg.de
2. Download SMPL version 1.0.0 for Python 2.7 (10 shape PCs)
3. Extract SMPL_python_v.1.0.0.zip and copy smpl/models/basicModel_f_lbs_10_207_0_v1.0.0.pkl in assets/SMPL

We use sequences from AMASS to test our model. To download the sequences follow these steps:

1. Sign in into https://amass.is.tue.mpg.de
2. Download the body data for the CMU motions (SMPL+H model)
3. Extract CMU.tar.bz2 in assets/CMU:

tar -C assets/ -xf ~/Downloads/CMU.tar.bz2 CMU/ 

To generate the deformed garment meshes for a given sequence:

python run_snug.py --motion assets/CMU/07/07_02_poses.npz --garment tshirt  --savedir tmp/07_02

The losses/physics.py contains the implementation of the energies that we use to train our models. The original paper had some typos:

• The lamé coefficients provided in the paper (μ=11.1 and λ=20.9) were already multiplied by the thickness of the fabric, so multiplying by the thickness again as indicated in Eq 9 would provide incorrect results. The correct value of the lamé coefficients are μ=2.36e4 and λ=4.44e4.

• Our implementation of the bending energy uses a scaling factor to account for the length of the edges and the area of the triangles. This scaling factor is computed as follows: $scale=\frac{l^2}{4(a_1 + a_2)}$, where $a_1$ and $a_2$ are the areas of the adjacent triangles and $l$ is the length of the shared edge.

The meshes can be rendered using the Blender scripts of our other projects:

• https://github.com/isantesteban/vto-learning-based-animation
• https://github.com/isantesteban/vto-garment-collisions