GAN Photo Editing

Shiva Peri - Assignment 5

Part 1: Inverting the Generator

In this part, we try to reconstruct a target image using various latent space encodings (Z, W, W+). We also compare various weightings of L1 pixel loss and Perceptual loss. In all of the images below, l1_wgt=10. Qualitatively, the best results occur when the weightings of the L1 and Perceptual losses are equal. Another note is that StyleGan drastically outperforms VanillaGan in terms of the quality of the reconstructions

perc_wgt
Target
VanillaGan Z
StyleGan Z
StyleGan W
StyleGan W+
0.01
0.1
1
10
100
1000

Part 2: Interpolate your Cats

In this part, we interpolate between reconstructions of two different cats. From the last part, we used the same weightings for L1 loss and Perceptual loss. Qualitatively, the W, W+ latent spaces produced better interpolation results than the Z space.

StyleGan Z
StyleGan W
StyleGan W+

Part 3: Scribble to Image

In this part, we use a masked scribble image to optimize a latent noise parameter into a minimized image. Ideally, the resultant cats retain the features of the input scribble. Qualitatively, the best results came from the W latent space. Overall, however, the results are not that great. The best results come from when the perc_wgt equals the l1_wgt.

Scribble
StyleGan W (perc_wgt=10)
StyleGan W (perc_wgt=100)
StyleGan W+ (perc_wgt=10)
StyleGan W+ (perc_wgt=100)