Empirical Variational Autoencoder
EVA swaps the VAE’s fixed Gaussian prior for latent priors predicted from the training data.
The paper says this can be added to VAEs with a single extra linear layer. The goal is to reduce the gap between the prior and posterior that hurts conventional VAE sampling. In experiments on image and sound synthesis, EVA is reported to match autoregressive diffusion baselines competitively while running inference much faster. Source: HF Daily Papers' note
The paper says this can be added to VAEs with a single extra linear layer. The goal is to reduce the gap between the prior and posterior that hurts conventional VAE sampling. In experiments on image and sound synthesis, EVA is reported to match autoregressive diffusion baselines competitively while running inference much faster. Source: HF Daily Papers' note
score 4