> ML_LITERATURE // KINGMA-WELLING-2013-AUTO-ENCODING-VARIATIONAL-BAYES_v1.0
Auto-Encoding Variational Bayes (VAE)
Diederik P. Kingma, Max Welling · International Conference on Learning Representations (ICLR) (2013)
foundational2013foundationalthirdPartyReproduced
Principal Contribution
Introduced the Auto-Encoding Variational Bayes (AEVB) algorithm and the reparameterization trick for continuous latent variables.
Operational Relevance
The perceptual autoencoder stage in Latent Diffusion Models (Stable Diffusion, Flux) relies directly on continuous VAE latent spaces.
Assumptions
- Intractable posterior distributions over latent variables can be approximated by a continuous encoder parameterized by a neural network
Limitations
- Reconstruction pixel loss leads to over-smoothed, blurry samples in high-dimensional image synthesis