Style-GAN-Prior

This project is an extension to the article Style Generator Inversion for Image Enhancement and Animation by Aviv Gabbay and Yedid Hoshen (GitHub). Two additional applications are introduced to this method: Image deblurring using a known blur kernel and image recoloring of gray images to RGB images.

Deblurring

In this task, images are deblurred using the style-GAN-prior and a known blur kernel.

Input image:

Gaussian Blur	Uniform Blur	Motion Blur
Gaussian Reconstruction	Uniform Reconstruction	Motion Reconstruction

Input image:

Gaussian Blur	Uniform Blur	Motion Blur
Gaussian Reconstruction	Uniform Reconstruction	Motion Reconstruction

Recoloring

In this task, gray images are recolored using the style-GAN-prior. The conversion of an RGB image to a gray image was performed using the formula: Y = 0.2126R + 0.7152G + 0.0722B

RGB Image	Gray Image	Recolored Image

Usage

Getting started

1. Clone the official StyleGAN repository.
2. Add the local StyleGAN project to PYTHONPATH.

Style Image Prior for Deblurring

Deblurring blurry images using a known blur kernel:

deblurring.py --blurred-imgs-dir <blurred-input-dir> --deblurred-imgs-dir <deblurred-output-dir>
    --latents-dir <output-latents-dir> --blur-kernel <blur-method>
    [--original-imgs-dir <originl-imgs-dir>]
    [--blurred-img-size BLURRED_IMG_HEIGHT BLURRED_IMG_WIDTH]
    [--learning-rate LEARNING_RATE]
    [--total-iterations TOTAL_ITERATIONS]

Notes:

1. The optional values for the blur-kernel flag are: gaussian/uniform/motion-vertical/motion-horizontal.
2. If the original-imgs-dir is specified, the chosen blur-kernel will be first applied to all images in this dir and the blurred images will be saved in the blurred-imgs-dir.
3. The default input image size is 256X256X3. This can be changed using the --blurred-img-size flag.

Style Image Prior for Image Recoloring:

Given a gray face image, reconstruct an RGB image:

coloring.py --gray-imgs-dir <gray-input-dir> --colored-imgs-dir <colored-output-dir>
    --latents-dir <output-latents-dir>
    [--original-imgs-dir <originl-imgs-dir>]
    [--img-size GRAY_IMG_HEIGHT GRAY_IMG_WIDTH]
    [--learning-rate LEARNING_RATE]
    [--total-iterations TOTAL_ITERATIONS]

Notes:

1. If the original-imgs-dir is specified, all images in this dir will be first converted to gray images using the formula Y = 0.2126R + 0.7152G + 0.0722B and saved in the gray-imgs-dir.
2. The default input image size is 256X256X3. This can be changed using the --img-size flag.