Variational Autoencoder on FashionMNIST

This project implements a basic Variational Autoencoder (VAE) to reconstruct and sample FashionMNIST images, based on the original concept from Kingma & Welling (2013). The goal was primarily to experiment with VAEs and evaluate reconstruction quality and sample realism using the FID score.

⚠️ Note: This implementation is a minimal version and not optimized. It can be improved by tuning hyperparameters (e.g. latent_dim, β) or using more advanced variants such as:

• β-VAE (Higgins et al.)
• VQ-VAE (van den Oord et al.)
• VAE-GAN
• Diffusion-based decoders