Simple Variational Diffusion Models

Note! This is still work in progress and may contain errors. If you lke to contribute, please leave a pull request.

A simple (~700 line) PyTorch implementation of the "Variational Diffusion Models" paper by Kingma et al. Much of it was inspired by the lucidrains and revsic repositories.

This repository is documented on my blog post. If any part of either the code or the blog is not well-documented. Please let me know!

Please also see references below.

Requirements

• Python 3.9
• CUDA 11.5

Running

1. git clone https://github.com/DavidRuhe/simple-variational-diffusion-models.git
2. cd simple-variational-diffusion-models
3. source activate_py3.sh
4. python main.py

To Do

1. Network adjustments reported in Appendix.
2. Fourier features.
3. Reproduce paper results.

References

ArXiV

• Kingma et al., 2021
• Ho et al., 2020
• Song et al., 2021
• Rombach et al., 2021

GitHub

• revsic/jax-variational-diffwave
• Lucidrains/denoising-diffusion-pytorch