cd
to test folder)Download the testing data from here.
If you want to test on the augmented testing images, download the images from here and put it to the testing data folder downloaded above, change the image folder here to _ma_img
.
Download pretrained checkpoints from here. The checkpoint trained without augmentation is better for testing set in the VITON. But the checkpoint trained with augmentation is more robust for in-the-wild images.
python test.py --name demo --resize_or_crop None --batchSize 1 --gpu_ids 0 --warp_checkpoint your_path_to_the_down_loaded_ckp/PFAFN_warp_epoch_101.pth --gen_checkpoint your_path_to_the_down_loaded_ckp/PFAFN_gen_epoch_101.pth --dataroot 'your_path_to_the_downloaded_test_data'
cd
to the train folder)For VITON dataset, download the training data from VITON_train and put the folder VITON_traindata
under the foldertrain/dataset
For perceptual loss computation, download the vgg checkpoint from VGG_Model and put vgg19-dcbb9e9d.pth
under the folder train/models
.
For other custom dataset, please generate the training data folder with the same structure as VITON_traindata
.
More specifically, you need to prepare human parsing, pose (18 key points, saved in .json file) and densepose (a heatmap, different body region has different value).
For human parsing, you can use Human parser.
For pose, you can use Openpose.
For dense pose, you can use Dense pose.
Note: To train the model on your own dataset, you may need to decrease the learning rate to 0.00001.
Note: if you want to train with augmentation, check here, here, here and here in the code.
sh scripts/train_PBAFN_stage1_fs.sh
sh scripts/train_PBAFN_e2e_fs.sh
sh scripts/train_PFAFN_stage1_fs.sh
sh scripts/train_PFAFN_e2e_fs.sh
If you find this repo helpful, please consider citing:
@inproceedings{he2022fs_vton,
title={Style-Based Global Appearance Flow for Virtual Try-On},
author={He, Sen and Song, Yi-Zhe and Xiang, Tao},
booktitle={CVPR},
year={2022}
}
This repository is based on PF-AFN, where we replaced the tensor correlation based flow estimation with our proposed style-based flow estimation.