CelebA report - CNN layers



Reconstructions

fancy-sweep-6

astral-sweep-5

eager-sweep-4

crisp-sweep-3

lemon-sweep-2

autumn-sweep-1

Step

Run set

6



encoded_dim = 64

Comments

Reconstructions look more accurate here, features such as gender, skin tone, hair color, face orientation seems to be captured
Some generations seems to be acceptable and coherent with the conditioned signal, but overall the space doesn't seem to be well regularized