Project title: Basic Diffusion Model for Simple Geometric Shape Generation

Overview

This repository contains a basic conceptual demonstration of a Denoising Diffusion Probabilistc Model (DDPM) implemented using PyTorch. The proroject focuses on generating two simple 2D geometric shapes (rectangles, circles) from random noise, showcasing a foundational understanding of diffusion-based generative models. This work serves as a practical exploration of AI methods relevant to many mainstream generative models such as Stable diffusion and DALL-E

Project Goals

The primary objectives of this project were to:

• Generate a synthetic dataset of simple geometric shapes using Python's OpenCV and NumPy.
• Implement a simplified DDPM from scratch in PyTorch, covering the forward (noising) and reverse (denoising) processes.
• Develop a U-Net architecture as the noise prediction network, incorporating timestep conditioning.
• Train the diffusion model on the generated dataset.
• Demonstrate image generation by sampling from pure noise and iteratively denoising.
• Document the process and findings in a clear, concise Jupyter Notebook.

Technical Details

• Dataset Generation: A custom script generates a dataset of 64x64 grayscale images featuring cicles and squares that are placed randomly within certain boundaries.
• Diffusion Process:
  • Forward Diffusion:
  • Backward Diffusion:
• U-Net:
  • A U-Net architecture is used as the backbone for predicting the noise at each timestep
  • It features a encoder-decoder structure with skip connections to preserve fine details.
  • Timestep Conditioning: The model is conditioned on the timestep by passing the timestep t through a small MLP and injecting it into the U-Net's bottleneck layer by addition, which allows the model to adapt the predicted noise depending on the timestep.
• Training: The model is trained by minimizing the Mean Squared Error (MSE) between the predicted noise and the noise added during the forward pass.

Results and Discussion

The ´diffusion_patterns.ipynb´notebook provides a setp by step walkthrough of the implementaion, including visualizations of:

• Examples from the generated datasets
• The Forward Diffusion proces
• Generated samples from the trained diffusion model at the final timestep as well as some intermediate timesteps. While the generated images are simple and a bit blurry, the demonstrate the model's ability to learn and generate images similar to that of the treaining data.

How to Run

1. Clone the repository

   • git clone https://github.com/maruand/Basic-Diffusion-Model-Demo

   • cd DDPM-Simple-Patterns

2. Create and activate a virtual environment:

   • python3 -m venv venv

   • source venv/bin/activate # On Windows: .\venv\Scripts\activate

3. Install dependencies
   • pip install torch torchvision torchaudio
   • pip install numpy opencv-python matplotlib jupyter pillow
4. Launch Jupyter Notebook
5. Open diffusion_patterns.ipynb and run all cells. The notebook will generated the dataset, train model and display results

Future Work

• Try more complex synthetic datasets or small real-world datasets such as MNIST or FashionMNIST
• Implement conditional generation
• Explore different timestep embeddings