Assignment #5 - GAN Photo Editing

Emily Kim

This assignment seeks to experiment with the pretrained GANs. First we invert the pretrained generator to reconstruct images. Then we reconstruct the images from scribbles.

Part 1: Inverting the Generator

Experimenting with various combinations of the losses, generative model, and latent space

These first set of images were reconstructed with Vanilla gan with z.

Target image	model: vanilla, latent: z, perc_wgt: 0.1, l1_wgt: 0.9	model: vanilla, latent: z, perc_wgt: 0.5, l1_wgt: 0.5	model: vanilla, latent: z, perc_wgt: 1, l1_wgt: 0

With StyleGAN, the images look better.

Target image	model: stylegan, latent: w, perc_wgt: 0.1, l1_wgt: 0.9	model: stylegan, latent: w, perc_wgt: 0.5, l1_wgt: 0.5	model: stylegan, latent: w+, perc_wgt: 0.5, l1_wgt: 0.5

Notice that the image reconstructed from StyleGAN, Latent option W, perception weight 0.5 and L1 weight 0.5 has less distortion around the eye compared to the cats. Vanilla GAN seems to not work very well with the small eyes of the target cat, but with the StyleGAN this issue gets solved.

Part 2: Interpolation between two target images

Here we interpolate between the images that have been reconstructed from two different target images. Each row corresponds to the same set of two target images. We can see that the cat with the small eyes (black cat with yellow eyes) is not performing very well for all of the results. With the cats that look similar in their appearance (white-brown with blue eyes) the images look more convincing.

StyleGAN Z	StyleGAN W	StyleGAN W+

Part 3: Scribble to Image

This part of the project seeks to reconstruct a cat image from a scribble image. The reconstructed images often look very dark or show very little details. Latent option W gives the most realistic results out of all the scribble images.

Target scribble	Target Mask	StyleGAN W	StyleGAN W+