HW 5

Bells and Whistles Attempted: Interpolate Cats

Q1

Vary the L1 weight terms (left to right: 1, 0.1, 0, 1000, 100)

L1=0.01 seems to give the best results

Vary the Perceptual loss weight (left to right: 9, 1, 0.1, 0.01, 0.001)

Perc_weight = 10 seems to best

Vary the architecture (left to right: vanilla, stylegan)

stylegan gives more closer results to gt

Vary the latent space (left to right: z, w, w+)

w+ is able to capture the color details much better

all models run within 20 seconds

Q2

We notice that results with sparser masks work better compared to dense masks, this is probably because with denser masks, more pixels are forced to be similar to the sketch -- and since the sketch is unrealistic, its hard for the model to converge easily.

Q3

Varying the noise (sketch, 100, 250, 500)

More noise helps make the generated image farther away from the sketch.

Varying the strenght of classifier-free guidance (sketch, 10, 15, 30, 60)

More strength of classifier-guidance makes the cats more grumpy -- and as expected makes it more aligned to the prompts

More results

Bells and Whistles

Interpolation