Semantically Decomposing the Latent Spaces of Generative Adversarial Networks
- Comparison of discriminator architecture on MS-Celeba_1M
- Result of SDGAN(stack the feature map along channels)
- Result of SDGAN(stack in RGB)
-
Decompose latent space into identities and observations. Z = [Z_i, Z_o]
-
Pairwise training scheme(Generator share weights, Discriminator share weights with D_e, Siamese setup for learning generating a pair of images(same identity).