VAEs
← Back to Generative Models
Variational Autoencoders. Learn a compressed latent space representation of data. Encode input to a distribution in latent space, sample from it, decode back. Loss combines reconstruction quality and KL divergence (regularize latent space). Enable interpolation and controlled generation.
Related
- GANs (alternative: adversarial training)
- Diffusion Models (alternative: iterative denoising)