Diffusion Models Implementation (DDPM & CDM)

This repository contains several implementations of diffusion models, starting with a Denoising Diffusion Probabilistic Model (DDPM) and progressing to a conditional version (CDM). These exercises allow exploring image generation using a diffusion and denoising process.

Repository Content

1. DDPM Implementation and Sampling Process

Objective:

• Understand how probabilistic diffusion models (DDPM) work.
• Learn how images are generated by reversing a noise diffusion process.
• Implement the sampling process to progressively denoise an image from pure noise.

2. DDPM Implementation and Training Process

Objective:

• Train a DDPM model on image data.
• Implement the loss function and train the denoising network.
• Observe the improvement in image quality as the model progresses.

3. Training a Conditional Diffusion Model (CDM)

Objective:

• Extend the DDPM approach to allow for conditional image generation.
• Condition the model on additional information (e.g., class labels, textual descriptions, structured attributes).
• Experiment with a Conditional Denoising Diffusion Model (CDDM) for more precise control over generated images.

Trained models are saved in the models directory.