Empirical Variational Autoencoder
Empirical VAE learns autoregressive latent priors with one linear layer, matching diffusion quality at faster inference.
Empirical Variational Autoencoder is a generative framework for continuous, non-vector-quantized sequences. It keeps the VAE evidence lower bound but replaces the standard Gaussian prior with autoregressive latent priors learned from data via a single extra linear layer. Image and sound experiments show generation quality competitive with autoregressive diffusion baselines and substantially faster inference.
- Adds one linear layer to learn data-driven autoregressive latent priors
- Reduces the prior-posterior latent gap typical of standard VAEs
- Matches autoregressive diffusion quality with faster inference on image and sound
From the original article
1 min read at huggingface.co
We present Empirical Variational Autoencoder, a general generative framework for continuous-valued (i. e. , non-vector-quantized) sequences.
By Kaede Shiohara · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.