Announcement
Empirical Variational Autoencoder proposes sequential generation with one extra linear layer
Its introduction says the added layer predicts the next prior, without VQVAEs or diffusion models.
TLDR
The announcement introduces Empirical Variational Autoencoder (EVA) and claims it can revive variational autoencoders for sequential generation. It says EVA needs just one extra linear layer to predict the next prior, without VQVAEs or diffusion models.
Combined views
30.6K
2 Sources, first seen ago
