STGAN (CVPR 2019)

Tensorflow implementation of STGAN: A Unified Selective Transfer Network for Arbitrary Image Attribute Editing

Overall architecture of our STGAN. Taking the image above as an example, in the difference attribute vector $\mathbf{att}_\mathit{diff}$ ,

is set to 1, $Mouth\ Open$ is set to -1, and others are set to zeros. The outputs of $\mathit{D_{att}}$ and $\mathit{D_{adv}}$ are the scalar $\mathit{D_{adv}}(\mathit{G}(\mathbf{x}, \mathbf{att}_\mathit{diff}))$ and the vector $\mathit{D_{att}}(\mathit{G}(\mathbf{x}, \mathbf{att}_\mathit{diff}))$ , respectively

Exemplar Results

See results.md for more results
Facial attribute editing results

Facial attribute editing results on the CelebA dataset. The rows from top to down are results of IcGAN, FaderNet, AttGAN, StarGAN and STGAN.

High resolution ( $384\times384$ ) results of STGAN for facial attribute editing.
Image translation results

Results of season translation, the top two rows are $summer \rightarrow winter$ , and the bottom two rows are $winter \rightarrow summer$ .

Preparation

Prerequisites
- Tensorflow (r1.4 - r1.12 should work fine)
- Python 3.x with matplotlib, numpy and scipy
Dataset
- CelebA dataset (Find more details from the project page)
  - Images should be placed in DATAROOT/img_align_celeba/*.jpg
  - Attribute labels should be placed in DATAROOT/list_attr_celeba.txt
  - If google drive is unreachable, you can get the data from Baidu Cloud
- We follow the settings of AttGAN, kindly refer to AttGAN for more dataset preparation details
pre-trained model
- download the model from Google Drive or Baidu Cloud (4qeu), and unzip the files to ./output/,

Quick Start

Exemplar commands are listed here for a quick start.

Training

for 128x128 images
```
python train.py --experiment_name 128
```

for 384x384 images (please prepare data according to HD-CelebA)

python train.py --experiment_name 384 --img_size 384 --enc_dim 48 --dec_dim 48 --dis_dim 48 --dis_fc_dim 512 --n_sample 24 --use_cropped_img

Testing

Example of testing single attribute

python test.py --experiment_name 128 [--test_int 1.0]

Example of testing multiple attributes

python test.py --experiment_name 128 --test_atts Pale_Skin Male [--test_ints 1.0 1.0]

Example of attribute intensity control

python test.py --experiment_name 128 --test_slide --test_att Male [--test_int_min -1.0 --test_int_max 1.0 --n_slide 10]

The arguments in [] are optional with a default value.

NOTE:

You should give the path of the data by adding --dataroot DATAROOT;
You can specify which GPU to use by adding --gpu GPU, e.g., --gpu 0;
You can specify which image(s) to test by adding --img num (e.g., --img 182638, --img 200000 200001 200002), where the number should be no larger than 202599 and is suggested to be no smaller than 182638 as our test set starts at 182638.png.
You can modify the model by using following arguments
- --label: 'diff'(default) for difference attribute vector, 'target' for target attribute vector
- --stu_norm: 'none'(default), 'bn' or 'in' for adding no/batch/instance normalization in STUs
- --mode: 'wgan'(default), 'lsgan' or 'dcgan' for differenct GAN losses
- More arguments please refer to train.py

AttGAN

Train with AttGAN model by

python train.py --experiment_name attgan_128 --use_stu false --shortcut_layers 1 --inject_layers 1

Citation

If you find STGAN useful in your research work, please consider citing:

@InProceedings{liu2019stgan,
  title={STGAN: A Unified Selective Transfer Network for Arbitrary Image Attribute Editing},
  author={Liu, Ming and Ding, Yukang and Xia, Min and Liu, Xiao and Ding, Errui and Zuo, Wangmeng and Wen, Shilei},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2019}
}

Acknowledgement

The code is built upon AttGAN, thanks for their excellent work!

lishali / STGAN