MSAI 437 Deep Learning Final Project - Simple DDIM from scratch

This repository contains the implementation of a simple Denoising Diffusion Implicit Models (DDIM) for the MSAI 437 Deep Learning course at Northwestern University. The project aims to demonstrate the application of diffusion models in generating high-quality samples from a data distribution.

Overview

Denoising Diffusion Implicit Models (DDIM) are a class of generative models that iteratively convert noise into samples from the target distribution. This project implements a simplified version of DDIM to understand the underlying principles and to explore its capabilities in generating complex data distributions.

Model outputs

• Dataset can be found here
• Download trained model weights from here and place in ./trained_models

Installation

To set up the project environment, follow these steps:

# Clone the repository
git clone https://github.com/yashkhurana24/ddim
cd ddim

# Create a virtual environment (optional)
conda create -n ddim python=3.8
conda activate ddim

# Install the required dependencies
# Install PyTorch from the official website (compiled with CUDA)
pip install -r requirements.txt

To-do

• training
• add attention mechanism to unet architecture
• evaluation of trained model
• inference from trained model
• add trained models and model outputs
• train GAN on same data for performance comparison
• remove all dependencies except pytorch