Assignment #5 - Cats Photo Editing

Part 1: Inverting the Generator

Combinations of the losses

Changing the Perpetual Loss while keeping the L1 loss at 10

Target	Perpetual Loss = 0	Perpetual Loss = 0.01	Perpetual Loss = 0.1	Perpetual Loss = 1	Perpetual Loss = 10

Looking at this output the perpetual loss was fixed to 0.01 which gave the best results.

Different generative models including vanilla GAN, StyleGAN

Target	Vanilla GAN	StyleGAN

Looking at this outputs StylegGAN gave the best results.

Different latent space (latent code in z space, w space, and w+ space)

Target	z	w	w+

Looking at the results w+ gives us the best results.

Part 2: Scribble to Image

Below are some results with the given set of images and some drawn images.

Sketch (S)	Mask (M)	Output

As we can see here using sketches with sparse lines and shapes gives good results.

Part 3: Stable Diffusion

Varied Noise

Prompt	Input	Noise SD: 0.5	Noise SD: 0.7
Grumpy cat reimagined as a royal painting
Dog reimagined as a royal painting

Varied Guidance

Prompt	Input	Guidance: 5	Guidance: 20
Grumpy cat reimagined as a royal painting
Dog reimagined as a royal painting