Awesome

Virtual Try-On with Garment-Pose Keypoints Guided Inpainting

This codes repository provides the pytorch implementation of the KGI virtual try-on method proposed in ICCV23 paper Virtual Try-On with Garment-Pose Keypoints Guided Inpainting.

Experimental Environment

Please follow the steps below to build the environment and install the required packages.

conda create -n kgi python=3.8 -y
conda activate kgi
bash install_pkgs.sh

Data Preparation

1. The VITON-HD dataset could be downloaded from VITON-HD . Please place the dataset under the directory KGI/data/. The dataset contains the following content:

   Content	Comment
   agnostic-mask	not in use in KGI
   agnostic-v3.2	not in use in KGI
   cloth
   cloth-mask	not in use in KGI
   image
   image-densepose	not in use in KGI
   image-parse-agnostic-v3.2	not in use in KGI
   image-parse-v3
   openpose_img	not in use in KGI
   openpose_json

2. In addition to above content, some other preprocessed conditions are in use in KGI. The content are generated with the data preprocessing codes [WIP]. The preprocessed data could be downloaded, respectively.

   Content	Train	Test
   image-landmark-json	Google Drive	Google Drive
   cloth-landmark-json	Google Drive	Google Drive
   label	Google Drive	Google Drive
   parse	Google Drive	Google Drive
   parse_ag_full	Google Drive	Google Drive
   ag_mask	Google Drive	Google Drive
   skin_mask	Google Drive	Google Drive

   • Data Preprocessing [WIP]

3. Download the demo_paired_pairs.txt and demo_unpaired_pairs.txt under the directory KGI/data/zalando-hd-resized/ for in-training visualization.

4. The structure of processed dataset should be as below:

   • KGI/data/zalando-hd-resized/
     • test/
       • ag_mask/
       • cloth/
       • cloth-landmark-json/
       • image/
       • image-landmark-json/
       • image-parse-v3/
       • openpose_json/
       • parse/
       • parse_ag_full/
       • skin_mask/
       • label.json
     • train/
       • ...
     • demo_paired_pairs.txt
     • demo_unpaired_pairs.txt
     • test_pairs.txt
     • train_pairs.txt

Model Training

The model training of the KGI method consists of three steps: Training of the Keypoints Generator, Training of the Parse Generator, Training of the Semantic Conditioned Inpainting Model.

Keypoints Generator

• The Keypoints Generator is trained with the following scripts:

  cd codes_kg
  python3 train_kg.py
  

  During the training, the visualization of some validation samples will be saved under directory KGI/visualizations/two_graph_cs/. Below is an example of visualization results. The pretrianed checkpoints of Keypoints Generator could be downloaded from Google Drive and put under the directory KGI/checkpoints_pretrained/kg/.
• Since the parse generation is based on the estimated keypoints conditions, please generate the keypoints conditions with the following scripts before the training of Parse Generator:

  bash generate_kg_demo_paired.sh
  bash generate_kg_demo_unpaired.sh
  bash generate_kg_train.sh
  bash generate_kg_test_paired.sh
  bash generate_kg_test_unpaired.sh
  

  The generated keypoints conditions will be saved under the directory KGI/example/generate_kg/ and also could be downloaded from train_kg_conditions and test_kg_conditions. The files should be placed under the directories KGI/data/zalando-hd-resized/train/ and KGI/data/zalando-hd-resized/test/, respectively.

Parse Generator

• The Parse Generator is trained with the following scripts:

  cd codes_pg
  python3 train_pg.py
  

  During the training, the visualization of some validation samples will be saved under directory KGI/visualizations/parse_full/. Below is an example of visualization results. The pretrianed checkpoints of Parse Generator could be downloaded from Google Drive and put under the directory KGI/checkpoints_pretrained/pg/.
• After the training of the Parse Generator, the person image parse (estimated segmentation map) could be generated with the following scripts:

  bash generate_pg_demo_paired.sh
  bash generate_pg_demo_unpaired.sh
  bash generate_pg_test_paired.sh
  bash generate_pg_test_unpaired.sh
  

  The generated parse conditions will be saved under the directory KGI/example/generate_pg/ and also could be downloaded from test_pg_conditions. The files should be placed under the directory KGI/data/zalando-hd-resized/test/ for tps conditions and final results generation.

Semantic Conditioned Inpainting Model

• The Semantic Conditioned Inpainting Model is trained with the following scripts:

  cd codes_sdm
  python3 train_sdm.py
  

  The pretrianed checkpoints of Semantic Conditioned Inpainting Model could be downloaded from Google Drive and put under the directory KGI/checkpoints_pretrained/sci/ckpt_1024/.
• The tps conditions (recomposed person image and content keeping mask) could be generated with the following scripts:

  cd codes_tps
  bash generate_tps_demo_paired.sh
  bash generate_tps_demo_unpaired.sh
  bash generate_tps_test_paired.sh
  bash generate_tps_test_unpaired.sh
  

  The generated tps conditions will be saved under the directory KGI/example/generate_tps/ and also could be downloaded from test_tps_conditions. The files should be placed under the directory KGI/data/zalando-hd-resized/test/ for final results generation with semantic conditioned inpainting.
• With the trained Semantic Conditioned Inpainting Model and tps conditions, the final results could be generated with the following scripts:

  cd codes_sci
  bash generate_sci_demo_paired.sh
  bash generate_sci_demo_unpaired.sh
  bash generate_sci_test_paired.sh
  bash generate_sci_test_unpaired.sh
  

Demo with Pretrained Model

With the pretrained models, the final try-on results and the visualizations of the intermediate results could be generated with the following demo scripts:

python3 generate_demo.py

The final try-on results will be saved under KGI/example/generate_demo/final_results/ and the visualizations of the intermediate results will be saved under KGI/example/generate_demo/vis/. Below is an example of demo results.

Acknowledgement and Citations

• The implementation of Keypoints Generator is based on codes repo SemGCN.
• The implementation of Semantic Conditioned Inpainting Model is based on semantic-diffusion-model and RePaint.
• The implementation of datasets and dataloader is based on codes repo HR-VITON.
• If you find our work is useful, please use the following citation:

  @InProceedings{Li_2023_ICCV,
    author    = {Li, Zhi and Wei, Pengfei and Yin, Xiang and Ma, Zejun and Kot, Alex C.},
    title     = {Virtual Try-On with Pose-Garment Keypoints Guided Inpainting},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {22788-22797}
  }