Variational AutoEncoder

Table of Contents

• Variational AutoEncoder
  • Introduction
  • Performance
  • Checkpoints
  • License

Introduction

This is an Pytorch implementation of Variational AutoEncoder, model structure was inspired by the one used with Stable-Diffusion.

Performance

The following is the result of training on the ImageNet-1K dataset with 256x256 resolution and 64 latent size for 55100 steps with batch size of 32. The final L1 loss is 0.015656.

Checkpoints

Checkpoints are in the sub-directory checkpoints, and are in the format of DATASET_IMAGESIZE_LATENTSIZE_STEPS-BATCHSIZE_LOSS.pt

License

vae is distributed under the terms of the MIT license.