The figures below show the architecture of the model from the data preprocessing step to the main HR-VITON model we used in this project.
We train and evaluate our model using the dataset from High-Resolution Virtual Try-On with Misalignment and Occlusion-Handled Conditions, the original dataset was from VITON-HD: High-Resolution Virtual Try-On via Misalignment-Aware Normalization.
To download the datasets, please check the links below:
After you download the dataset, create a ./data
folder and put it under.
If you want to know more about how we preprocess the data, please check the Preprocessing.md
Here are the download links for each model checkpoint:
- Author's try-on condition generator: link
- Our retrain 30000 steps try-on condition generator: link
- Author's try-on image generator: link
- We assume that you have obtained all the checkpoints and stored them in
./eval_models/weights/v0.1
.
We've built a web app demo, please check the following notebook for detailed instructions
python train_condition.py --name test --cuda {True} --gpu_ids {gpu_ids} --dataroot {dataroot_path} --datamode train --data_list {path_to_trainpairs.txt} --keep_step 300000 --Ddownx2 --Ddropout --lasttvonly --interflowloss --occlusion
python train_generator.py --cuda {True} --name test -b 4 -j 8 --gpu_ids {gpu_ids} --fp16 --tocg_checkpoint {condition generator ckpt path} --occlusion --dataroot {data_path}
To use "--fp16" option, you should install the apex library.
- Tran Ngoc Xuan Tin ([email protected])
- Duong Vien Thach ([email protected])
- https://github.com/sangyun884/HR-VITON
- https://github.com/shadow2496/VITON-HD
- https://github.com/plemeri/transparent-background
- https://github.com/CMU-Perceptual-Computing-Lab/openpose
- https://github.com/facebookresearch/detectron2/tree/main/projects/DensePose
- https://github.com/Engineering-Course/CIHP_PGN