This repository implements a variational auto-encoder (VAE) for MNIST based on the Auto-Encoding Variational Bayes paper. It uses TensorFlow to train a generative model that learns a latent representation of digit images. The implementation supports image reproduction, denoising with salt-and-pepper noise, and visualization of learned latent manifolds. The model is trained with configurable latent space dimensionality (2D to 20D).