Paper | Supp

Discovering Interpretable Latent Space Directions of GANs Beyond Binary Attributes.

Huiting Yang, Liangyu Chai, Qiang Wen, Shuang Zhao, Zixun Sun, Shengfeng He

In CVPR 2021

• Linux
• Python 3
• NVIDIA GPU + CUDA CuDNN

• Clone this repo:

git clone https://github.com/BERYLSHEEP/AdvStyle.git

• Download resources:
  • Please download the pretrained FFHQ StyleGAN and anime StyleGAN weights
  • Move the weights to models/pretrain/pytorch

The following commands are examples of testing the learned direction:

# stylegan ffhq
python new_demo.py manipulate_test supermodel \
				--gan_model stylegan_ffhq --resolution 1024 --latent_type z
	
# stylegan anime
python new_demo.py manipulate_test maruko \
				--gan_model stylegan_anime --resolution 512 --latent_type z

if you want to manipulate multi attributes simultaneously, you can list all the attributes on the command as follow:

# multi attribute manipulation
python new_demo.py manipulate_test blonde,open_mouth \
				--gan_model stylegan_anime --resolution 512 --latent_type z

If you want to specific the manipulated latent code, you can set the --noise_path option:

# specific latent code
python new_demo.py manipulate_test maruko \
				--gan_model stylegan_anime --resolution 512 --latent_type z \
				--noise_path ./noise/maruko/2.npy

Results are saved to result/{attribute}

The attribute names are the file names in the boundaries directory.

• Prerequisites: Flask

  pip install flask
  pip install --upgrade flask_nav
  pip install --upgrade flask_bootstrap

• Running

  # only for stylegan_anime
  # the input latent code is belong to latent space z [1, 512]
  python app.py --latent_code noise/maruko/16.npy --composing_type z
  # the input latent code is belong to latent space w+ [1, 16, 512]
  python app.py --latent_code noise/38_w+.npy --composint_type w+

• Instructions

  • Only one attribute can be edited at once time.
  • Multi attributes edit examples:
    • Click the corresponding Edit Button for "Pink Hair", then move the corresponding toolbar to the desired step.
    • Click the corresponding Edit Button for "Open Mouth", then move the corresponding toolbar to the desired step.
    • ....

• The pretrained StyleGAN weights are from official StyleGAN and Gwern
• Using the pytorch implementation of StyleGAN from [InterFaceGAN](Pytorch implementation of official model)

If you use this code for your research, please cite our paper:

@InProceedings{Yang_2021_CVPR,
    author    = {Yang, Huiting and Chai, Liangyu and Wen, Qiang and Zhao, Shuang and Sun, Zixun and He, Shengfeng},
    title     = {Discovering Interpretable Latent Space Directions of GANs Beyond Binary Attributes},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {12177-12185}
}