HF Paper EVA Swaps VAE's Standard-Gaussian Prior for Self-Predicted Autoregressive Priors, Matches Diffusion Quality Faster

A paper posted to Hugging Face Oct 6 proposes Empirical Variational Autoencoder (EVA), which replaces the standard-Gaussian prior constraint in VAEs with autoregressive latent priors learned from training data — implemented as a single extra linear layer. The authors argue EVA closes the prior-posterior gap that hurts VAE ancestral sampling, hitting generation quality competitive with autoregressive diffusion baselines on ImageNet-256 images and audio synthesis at much faster inference.