AI engineering glossary · Generative media multimodal
What is Beta-VAE?
Beta-VAE: A VAE variant that multiplies the KL term in the objective by a factor beta greater than 1. This pressure encourages more independent, interpretable latent dimensions at the cost of blurrier reconstructions.
Beta-VAE: an example
With beta = 4, one latent dimension of a face model may control only smile while another controls only head angle.
Related terms
- Variational autoencoder
- Disentangled representation
- Evidence lower bound
More Generative media multimodal terms
- BLIP
- Classifier-free guidance
- CLIP score
- Codebook
- Conditional GAN
- Consistency model
- Content provenance
- ControlNet
Where you learn it
Beta-VAE is part of the AI Engineering Program, which teaches it with notes, a hands-on lab and an exam. The first lesson is free.