Implement diffusion denoising models from scratch

• run setup.sh to setup your environment
• run code in download.ipynb to gather ~1000 images from miyazaki movies
• use training.ipynb to train a model and generate sample imagery

References:

The basic functions for training are taken from the original ddpm paper. The code here is commented with references to the specific parts the paper's algorithm.

For the Unet architecture, two are tried here

• this: https://github.com/dome272/Diffusion-Models-pytorch/blob/main/modules.py
• and this: https://huggingface.co/blog/annotated-diffusion

Both are similar and gave similar results but ultimately the former was chosen since it's a bit simpler/smaller and easier to play with

In addition, some of the image transforms are taken from those sources as well.