Diffusion from Scratch

This repo contains a guide to implementing diffusion models from scratch.

Chapters

1. Forward Process

   We begin by building a simple forward process that adds Gaussian noise to an image. This is how we'll generate our input data for training our diffusion models.

2. Reverse Process / Training

   We implement the loss function, and training loop, and train a very simple model.

3. Sampling / Inference

   We show how to sample from our trained model to generate images.

4. MNIST

   We train a model on the MNIST dataset.

5. CIFAR10

   We train a model on the CIFAR10 dataset.

6. Conditional Models

   We show how to train conditional models, allowing us to generate samples from a specific class.

7. CelebA

   We train a model on the CelebA dataset.