Full-body Virtual Try-On using Top and Bottom Garments with Wearing Style Control

Official implementation for "Full-body Virtual Try-On using Top and Bottom Garments with Wearing Style Control" published in Computer Vision and Image Understanding in 2024.

• Author: Soonchan Park and Jinah Park
• Paper link: https://www.sciencedirect.com/science/article/pii/S1077314224003400
• Supplementary material: https://ars.els-cdn.com/content/image/1-s2.0-S1077314224003400-mmc1.pdf

Dataset : Fashion-TB

Fashion-TB dataset consists of 37,492 wearing pairs between garments and fasion model. The fashion model can has multiple indices to point top and bottom garments what he/she is wearing. In dress case, such a model has only one index for top garment.

Installation

The project have been tested with the following environment:

• PyTorch 1.13.1
• Torchvision 0.13.1
• Python 3.8
• CUDA 11.6

Docker

The folder "docker" has Dockerfile to set docker images for running WGF-VITON.

cd docker
docker build . -t {docker_image_name}

Anaconda

conda create -n {name} python=3.8 anaconda
conda activate {name}
conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia
pip install cupy

Downloads

You can download dataset and pre-trained model via links below. Please note license policy of this work.

• pre-trained model of WGF-VITON and Fashion-TB: https://drive.google.com/drive/folders/12PoisAUqyyA4AOWjXMdFJlR4OrI9DrcO?usp=sharing
• released version of WGF-VITON has 6.72 FID in the unpaired dataset. It is achieved by a longer training without any modification.

Training

CUDA_VISIBLE_DEVICES=0,1 python train.py --dataroot {data_path} --gpu_ids 0,1 --batch-size 4 --name {project_name} --workers 6 --keep_step 100000 --decay_step 100000 --shuffle --fine_height 512 --fine_width 384 --num_D 2 --lr_D 0.0002

Testing

CUDA_VISIBLE_DEVICES=2 python test.py --name {project_name} --batch-size 8 --workers 1 --checkpoint {checkpoint_path}/step_G_200000.pth --wearing test_unpair_mild.json --dataroot {data_path} --gpu_ids 0

After inference dataset using the command above, you can evaluate the results in FID (https://github.com/mseitzer/pytorch-fid)

License

All material is made available under Creative Commons BY-NC 4.0. You can use, redistribute, and adapt the material for non-commercial purposes, as long as you give appropriate credit by citing our paper and indicate any changes that you've made.

Citation

@article{park2024full,
  title={Full-body virtual try-on using top and bottom garments with wearing style control},
  author={Park, Soonchan and Park, Jinah},
  journal={Computer Vision and Image Understanding},
  pages={104259},
  year={2024},
  publisher={Elsevier}
}

Acknoledgements

We implemente the code for WGF-VITON based on PyTorch implementation of CP-VTON, SPADE, and HR-VITON.