A PyTorch implementation of Denoising Diffusion Probabilistic Models (DDPM) trained on the MNIST dataset. This implementation is based on the assignments from 6.S978 Deep Generative Models (MIT EECS, Fall 2024).

This implementation includes:

• A U-Net architecture with residual connections and time embeddings
• DDPM training and sampling procedures
• Visualization of generated samples and intermediate diffusion steps

This implementation is a basic example and not a state-of-the-art model. Areas for improvement include:

• Limited network capacity (using basic U-Net architecture)
• No advanced sampling techniques like DDIM
• Basic time embeddings without attention mechanisms
• No conditioning capabilities
• Limited hyperparameter tuning
• Basic learning rate schedule

• 6.S978 Deep Generative Models, pset4 (MIT EECS, Fall 2024)