AttGAN

Tensorflow implementation of AttGAN - Arbitrary Facial Attribute Editing: Only Change What You Want

Exemplar Results

• See results.md for more results, we try higher resolution and more attributes (all 40 attributes!!!) here

• Inverting 13 attributes respectively

  from left to right: Input, Reconstruction, Bald, Bangs, Black_Hair, Blond_Hair, Brown_Hair, Bushy_Eyebrows, Eyeglasses, Male, Mouth_Slightly_Open, Mustache, No_Beard, Pale_Skin, Young

• Comparisons with VAE/GAN and IcGAN on inverting single attribute

• Comparisons with VAE/GAN and IcGAN on simultaneously inverting multiple attributes

Usage

• Prerequisites

• Tensorflow 1.7 or 1.8

• Python 2.7 or 3.6

• Dataset

• the images of img_align_celeba.zip are low resolution and uncropped, higher resolution and cropped images are available here

• the high quality data should be placed in ./data/img_crop_celeba/*.jpg

• Images should be placed in ./data/img_align_celeba/*.jpg

• Attribute labels should be placed in ./data/list_attr_celeba.txt

• the above links might be inaccessible, the alternative is

• https://pan.baidu.com/s/1eSNpdRG#list/path=%2FCelebA%2FAnno&parentPath=%2F or

• https://drive.google.com/drive/folders/0B7EVK8r0v71pOC0wOVZlQnFfaGs

• https://pan.baidu.com/s/1eSNpdRG#list/path=%2FCelebA%2FImg or

• https://drive.google.com/drive/folders/0B7EVK8r0v71pTUZsaXdaSnZBZzg

• img_align_celeba.zip

• list_attr_celeba.txt

• Celeba dataset

• HD-Celeba (optional)

• Well-trained models: download the models you need and unzip the files to ./output/ as below,

  output
   128_shortcut1_inject1_none
   384_shortcut1_inject1_none_hd

• Examples of training

• for 128x128 images

  CUDA_VISIBLE_DEVICES=0 python train.py --img_size 128 --shortcut_layers 1 --inject_layers 1 --experiment_name 128_shortcut1_inject1_none

• for 384x384 images

  CUDA_VISIBLE_DEVICES=0 python train.py --img_size 384 --enc_dim 48 --dec_dim 48 --dis_dim 48 --dis_fc_dim 512 --shortcut_layers 1 --inject_layers 1 --n_sample 24 --experiment_name 384_shortcut1_inject1_none

• for 384x384 HD images (need HD-Celeba)

  CUDA_VISIBLE_DEVICES=0 python train.py --img_size 384 --enc_dim 48 --dec_dim 48 --dis_dim 48 --dis_fc_dim 512 --shortcut_layers 1 --inject_layers 1 --n_sample 24 --use_cropped_img --experiment_name 384_shortcut1_inject1_none_hd

• see examples.md for more examples

• training

• tensorboard for loss visualization

  CUDA_VISIBLE_DEVICES='' tensorboard --logdir ./output/128_shortcut1_inject1_none/summaries --port 6006

• Example of testing single attribute

  CUDA_VISIBLE_DEVICES=0 python test.py --experiment_name 128_shortcut1_inject1_none --test_int 1.0

• Example of testing multiple attributes

  CUDA_VISIBLE_DEVICES=0 python test_multi.py --experiment_name 128_shortcut1_inject1_none --test_atts Pale_Skin Male --test_ints 0.5 0.5

• Example of attribute intensity control

  CUDA_VISIBLE_DEVICES=0 python test_slide.py --experiment_name 128_shortcut1_inject1_none --test_att Male --test_int_min -1.0 --test_int_max 1.0 --n_slide 10

Citation

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

@article{he2017arbitrary,
  title={Arbitrary Facial Attribute Editing: Only Change What You Want},
  author={He, Zhenliang and Zuo, Wangmeng and Kan, Meina and Shan, Shiguang and Chen, Xilin},
  journal={arXiv preprint arXiv:1711.10678},
  year={2017}
}