Generative Models

Variational Autoencoder (VAE)

Autoencoder Variational Autoencoder

A Variational Autoencoder (VAE) is a generative model that learns to encode input data into a probabilistic latent space and reconstruct it through a decoder. Unlike traditional autoencoders, VAEs model uncertainty by learning a distribution over latent variables, enabling smooth sampling and interpolation.

D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” International Conference on Learning Representations (ICLR), 2014.

Name

Model

Input Shape

VAE

VAE

\((N,C,H,W)\)

DDPM

Diffusion

A Denoising Diffusion Probabilistic Model (DDPM) is a generative model that learns to generate data by reversing a gradual noising process. It adds Gaussian noise over several timesteps and trains a neural network to denoise and recover the original data distribution through a Markovian process.

J. Ho, A. Jain, and P. Abbeel, “Denoising Diffusion Probabilistic Models,” Advances in Neural Information Processing Systems (NeurIPS), 2020.

Name

Model

Input Shape

Parameter Count

DDPM

DDPM

\((N,C,H,W)\)

20,907,649 (Default)

NCSN

Diffusion Score-Based Diffusion

A Noise Conditional Score Network (NCSN) is a score-based generative model trained to predict the score of noise-perturbed data across multiple noise levels, and it generates samples using annealed Langevin dynamics over a descending noise schedule.

Song, Yang, and Stefano Ermon. “Generative Modeling by Estimating Gradients of the Data Distribution.” Advances in Neural Information Processing Systems (NeurIPS), 2019, arXiv:1907.05600.

Name

Model

Input Shape

Parameter Count

NCSN

NCSN

\((N,C,H,W)\)

12,471,555 (Default)

To be implemented…🔮