Figure: Multi-code GAN prior facilitates many image processing applications using the reconstruction from fixed GAN models.

In this work, we propose a new inversion approach to applying well-trained GANs as effective prior to a variety of image processing tasks, such as image colorization, super-resolution, image inpainting, and semantic manipulation.

Image Processing Using Multi-Code GAN Prior
Jinjin Gu, Yujun Shen, Bolei Zhou
Computer Vision and Pattern Recognition (CVPR), 2020

[Paper] [Project Page]

Install dependencies using the following code before performing Multi-Code GAN Inversion.

python -m pip install -r requirements.txt

In this work, we use the well-trained GAN models as prior, including PGGAN and StyleGAN. Pytorch version models are borrowed from HiGAN. See here for more details.

As both PGGAN and StyleGAN use aligned face for GAN training, all faces used in this repo are pre-aligned. The alignment method can be found at stylegan-encoder.

With a given GAN model and a target image, you can invert the image to multiple latent codes by running

python multi_code_inversion.py
       --gan_model pggan_bedroom
       --target_images ./examples/gan_inversion/bedroom
       --outputs ./gan_inversion_bedroom
       --composing_layer 8
       --z_number 20

For image colorization task, run

python colorization.py
       --gan_model pggan_bedroom
       --target_images ./examples/colorization/bedroom
       --outputs ./colorization
       --composing_layer 6
       --z_number 20

For image inpainting task (inpainting mask should be known in advance), run

python inpainting.py
       --gan_model pggan_churchoutdoor
       --target_images ./examples/inpainting/church
       --outputs ./inpainting
       --mask ./examples/masks/mask-1.png
       --composing_layer 4
       --z_number 30

For image super-resolution task (SR factor should be known in advance), run

python super_resolution.py
       --gan_model pggan_celebahq
       --target_images ./examples/superresolution
       --outputs ./SR_face
       --factor 16
       --composing_layer 6
       --z_number 20

We achieve semantic face editing together with InterfaceGAN. Please refer to this repo to see how to train semantic boundaries in the latent space as well as how to achieve face manipulation by varying the latent code.

In this project, you can simply run

python face_semantic_editing.py
       --gan_model pggan_celebahq
       --target_images ./examples/face
       --outputs ./face_manipulation
       --attribute_name gender
       --composing_layer 6
       --z_number 30

@inproceedings{gu2020image,
  title     = {Image Processing Using Multi-Code GAN Prior},
  author    = {Gu, Jinjin and Shen, Yujun and Zhou, Bolei},
  booktitle = {CVPR},
  year      = {2020}
}