Diffusion Models

This project implements DDM paper for image generation using the CIFAR10 dataset. This is an educational project created to understand and learn about diffusion models, which are a class of generative models that learn to generate data through a gradual diffusion process.

Example of generated image

Architecture

• U-Net Architecture: Implementation of a U-Net network for the denoising process
• Sinusoidal Embedding: Temporal sinusoidal encoding for the diffusion process

Installation

1. Clone the repository:

git clone <repository-url>
cd Diffusion-models

2. Install dependencies:

pip install torch torchvision datasets matplotlib numpy

Usage

Training

To train the diffusion model:

python train.py

Image Generation

To generate images with the trained model:

python generate.py imgs/image.png

Configuration

You can modify parameters in config.py:

NUM_TIMESTEPS = 200      # Diffusion steps
EPOCHS = 40             # Training epochs
BATCH_SIZE = 8          # Batch size
BETA_MIN = 1e-4         # Minimum beta for schedule
BETA_MAX = 0.02         # Maximum beta for schedule
EMBEDDING_DIM = 256     # Temporal embedding dimension
INPUT_SHAPE = (3, 32, 32)  # Input shape (CIFAR10)