4639475d6782a08c1e964f9a4329a254-Supplemental.pdf

Additionally,weprovide21 the architectures used in each domain for our baselines: StyleGAN2 in Table 13 each domain in22 Table 13 and VAE in Table 8. Despitethis,wewereabletorun35 the models on toy datasets and found that these default hyperparameters performed the best. We utilize the Fourier embedding from [4]toembed coordinates. Latentencodings from image and audio modalities are added together.