VAE 0.05 β-VAE 0.18 β-TCVAE 0.30 DIP-VAE-II0.12 C-VITAE 0.38 Table1: dSprites Figure1: CelebA

Figure 18 illustrates how our model is able to capture geometrical face information such as the shape of the face. R2: Indeed, it should be possible to exchange the VAE framework with a WAE, by essentially changing the loss32 function. Wechoose explicitlytoworkondatasets where wehaveaccess togenerative47 factors, such that our results could be quantitatively evaluated.