ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns

This is the repo for ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns.

Setup:

Download checkpoints from here, and put *.pth at ./checkpoints.

Download and extract the SMPL model from http://smplify.is.tue.mpg.de/, and place basicModel_f_lbs_10_207_0_v1.0.0.pkl and basicModel_neutral_lbs_10_207_0_v1.0.0.pkl in the folder of ./smpl_pytorch.

The code is implemented with python 3.8, torch 2.0.1 and cuda 11.8 (other versions usually work as well).

Other dependencies include trimesh, pytorch3D, scipy.

Inference

For garment generation:

python infer_isp.py --which tee/pants/skirt --save_path tmp --save_name skirt --res 256 --idx_G 0

For layering inference:

python infer_layering.py

Fitting

For fitting ISP to 3D garment mesh in rest pose:

python fitting_3D_mesh.py --which tee/pants/skirt --save_path tmp --save_name skirt-fit --res 256

For fitting ISP to images:

python fitting_image.py

The example files are under ./extra-data/fitting-sample/, including the segmentation mask mask.png and the SMPL parameters mocap.pkl. We use Self-Correction-Human-Parsing to produce garment masks, and frankmocap to estimate SMPL parameters.

Citation

If you find our work useful, please cite it as:

@inproceedings{Li2023isp,
  author = {Li, Ren and Guillard, Benoit and Fua, Pascal},
  title = {{ISP: Multi-Layered Garment Draping with Implicit Sewing Patterns}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023}
}