Awesome

Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content, CVPR'20.

Rearranged code of CVPR 2020 paper 'Towards Photo-Realistic Virtual Try-On by Adaptively Generating↔Preserving Image Content' for open-sourcing. We rearrange the VITON dataset for easy access.

Notably, virtual try-on is a difficult research topic, and our solution is of course not perfect. Please refer to our failure cases and limitations before using this repo.

The code is not fully tested. If you meet any bugs or want to improve the system, please feel free to raise in the Issue and we can disscuss. For email request, please send to hanyang@ethz.ch

[Sample Try-on Video] [Checkpoints]

[Dataset_Test] Dataset_Train

[Paper]

Update

[2022-7-5] We collect the try-on results of several methods given a widely used test pair list in VITON dataset (from CP-VTON). Researchers can utilize it for fast baseline comparison.

CP-VTON+(CVPRW 2020), ACGPN(CVPR2020), DCTON(CVPR2021) and RT-VTON(CVPR 2022) are given in the results from left to right. [Datalink], [Test Pair List].
[2021-12-13] We remove the random dropout, and we use AdamW with weight decay to stablize training. clothes are pasted back before computing GAN loss and vgg loss. The light point artifacts are largely reduced. Code is Updated.
[2021-12-3] The light point artifacts seem to be caused by the variance of the imprecise human parsing when we rearrange the data for open-sourcing. We recommend to use the ATR model in https://github.com/PeikeLi/Self-Correction-Human-Parsing to get the human parsing with neck label to stablize training. To note that, face and neck part should be treated as non-target body part in mask inpainting and mask composition. With neck label, we can paste back the background before computing vgg loss and gan loss. The uncertainty of background might be another cause of the light point on the neck.
[2021-10-22] The light point artifacts would occur in current training results. This may be due to some version differences of our training codes when we rearranged them since we didn't observe same artifacts in our released checkpoints. It might be caused by the instablity in training the preservation (identical mapping) of clothes region in Content Fusion Module. Try to paste back the ground-truth clothes to the CFM results when calculating the VGG loss, Gan loss, Feature Matching loss (All except L1), since the above loss might degenerate the results when learning identical mapping. L1 loss can be applied to the reconstruction of clothes region to learn this identical mapping. This ISSUE addressed this problem.

Inference

python test.py

Note that the results of our pretrained model are only guaranteed in VITON dataset only, you should re-train the pipeline to get good results in other datasets.

Inference using colab

Thanks Levin for contributing the colab inference script.

Evaluation IS and SSIM

Note that The released checkpoints are different from what we used in the paper which generate better visual results but may have different (lower or higher) quantitative statistics. Same results of the paper can be reproduced by re-training with different training epochs.

The results for computing IS and SSIM are same-clothes reconstructed results.

The code defaultly generates random clothes-model pairs, so you need to modify ACGPN_inference/data/aligned_dataset.py to generate the reconstructed results.

Here, we also offer the reconstructed results on test set of VITON dataset by inferencing this github repo, [Precomputed Evaluation Results] The results here can be directly used to compute the IS and SSIM evalutations. You can get identical results using this github repo.

SSIM score

Use the pytorch SSIM repo. https://github.com/Po-Hsun-Su/pytorch-ssim
Normalize the image (img/255.0) and reshape correctly. If not normalized correctly, the results differ a lot.
Compute the score with window size = 11. The SSIM score should be 0.8664, which is a higher score than reported in paper since it is a better checkpoint.

IS score

Use the pytorch inception score repo. https://github.com/sbarratt/inception-score-pytorch
Normalize the images ((img/255.0)*2-1) and reshape correctly. Please strictly follow the procedure given in this repo.
Compute the score. The splits number also changes the results. We use splits number =1 to compute the results.
Note that the released checkpoints produce IS score 2.82, which is slightly lower (but still SOTA) than the paper since it is a different checkpoint with better SSIM performance.

The specific key points we choose to evaluate the try-on difficulty

We use the pose map to calculate the difficulty level of try-on. The key motivation behind this is the more complex the occlusions and layouts are in the clothing area, the harder it will be. And the formula is given below. Also, manual selection is involved to improve the difficulty partition.
Variations of the pose map predictions largely affect the absolute value of try-on complexity, so you may have different partition size using our reported separation values.
Relative ranking of complexity best depicts the complexity distribution. Try top 100 or bottom 100 and you can see the effectiveness of our criterion.

The formula to compute the difficulty of try-on reference image

where t is a certain key point, Mp' is the set of key point we take into consideration, and N is the size of the set.

Segmentation Label

0 -> Background
1 -> Hair
4 -> Upclothes
5 -> Left-shoe 
6 -> Right-shoe
7 -> Noise
8 -> Pants
9 -> Left_leg
10 -> Right_leg
11 -> Left_arm
12 -> Face
13 -> Right_arm

Sample images from different difficulty level

Sample Try-on Results

Limitations and Failure Cases

1. Large transformation of the semantic layout is hard to handle, partly ascribing to the agnostic input of fused segmentation. 2. The shape of the original clues is not completely removed. The same problem as VITON. 3. Very difficult pose is hard to handle. Better solution could be proposed.

Training Details

Due to some version differences of the code, and some updates for better quality, some implementation details may be different from the paper.

For better inference performance, model G and G2 should be trained with 200 epoches, while model G1 and U net should be trained with 20 epoches.

License

The use of this software is RESTRICTED to non-commercial research and educational purposes.

Citation

If you use our code or models or the offered baseline results in your research, please cite with:

@InProceedings{Yang_2020_CVPR,
author = {Yang, Han and Zhang, Ruimao and Guo, Xiaobao and Liu, Wei and Zuo, Wangmeng and Luo, Ping},
title = {Towards Photo-Realistic Virtual Try-On by Adaptively Generating-Preserving Image Content},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

@inproceedings{ge2021disentangled,
  title={Disentangled Cycle Consistency for Highly-realistic Virtual Try-On},
  author={Ge, Chongjian and Song, Yibing and Ge, Yuying and Yang, Han and Liu, Wei and Luo, Ping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16928--16937},
  year={2021}
}

@inproceedings{yang2022full,
title = {Full-Range Virtual Try-On With Recurrent Tri-Level Transform},
author = {Yang, Han and Yu, Xinrui and Liu, Ziwei},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages = {3460--3469}
year = {2022}
}

Dataset

VITON Dataset This dataset is presented in VITON, containing 19,000 image pairs, each of which includes a front-view woman image and a top clothing image. After removing the invalid image pairs, it yields 16,253 pairs, further splitting into a training set of 14,221 paris and a testing set of 2,032 pairs.