Course contentsShow
← Machine Learning and Deep Learning
Lesson 1539 of 3,53834. Diffusion ModelsFree lesson

DDPM Framework Overview

Understanding the complete denoising diffusion probabilistic model architecture and its training-inference cycle.

DDPM Framework Overview

What you'll learn: How all the pieces of Denoising Diffusion Probabilistic Models (DDPM) fit together into a complete training and inference system.

The Complete Picture

Now that you understand forward diffusion, reverse denoising, and the intuition behind each component, let's see how DDPM unifies everything into a single, elegant framework.

DDPM is essentially a two-phase dance:

  1. Forward diffusion (fixed, no learning): Gradually add Gaussian noise to data over T timesteps using a predefined variance schedule, until the data becomes pure noise.

  2. Reverse diffusion (learned): Train a neural network to reverse this process—learning to predict and remove noise at each timestep, conditioned on the current timestep number.

The Training-Inference Cycle

Training: You sample real data, randomly pick a timestep t, add the corresponding amount of noise (using the forward process closed form), then train your network to predict the noise that was added. The network takes the noisy image and timestep t as input. Loss is typically mean squared error between predicted and actual noise.

Inference: Start with pure Gaussian noise. For t = T down to t = 1, use your trained network to predict the noise at that timestep, subtract it (with careful variance handling), and gradually reveal a clean sample. This Markov chain walks backward through the learned reverse process.

Why It Works

The framework cleverly sidesteps the complexity of directly modeling high-dimensional data distributions. Instead, it learns a sequence of simple denoising steps—each one slightly easier than the last.

Key Takeaway: DDPM combines a fixed forward noising schedule with a learned reverse denoising network, training on noise prediction and sampling via iterative denoising from pure noise to data.