DDPM Explained: Denoising Diffusion Probabilistic Models

Added:

Flow Matching and Rectified Flows

While Denoising Diffusion Probabilistic Models (DDPMs) generate high-quality images, they suffer from slow sampling speeds and complex mathematical formulations due to their curved, stochastic trajectory through noise space. Flow Matching (and Rectified Flows) has emerged as a major alternative paradigm. Instead of relying on a complex, step-by-step probabilistic diffusion process, Flow Matching trains models using straight-line paths (vector fields) between noise and data. This formulation simplifies the underlying mathematics, enables faster training, and allows for much quicker, deterministic sampling with fewer steps. By straightening the generation trajectory, Flow Matching directly addresses the computational inefficiencies and mathematical complexity inherent in traditional DDPMs.

Fundamentals of probability theory, specifically multivariate Gaussian distributions, joint/conditional probabilities, and Bayes' theorem.
The concept of Markov Chains, as the diffusion forward and reverse processes are modeled as sequences of state transitions.
Variational Autoencoders (VAEs) and the Variational Lower Bound (ELBO), which form the direct mathematical foundation for the DDPM loss function.
The U-Net neural network architecture, which is the standard backbone used in DDPMs to predict noise at each timestep.
Denoising Diffusion Implicit Models (DDIMs) to understand how to dramatically accelerate the sampling/generation process.
Classifier-Free Guidance (CFG) and conditioning techniques, which enable text-to-image generation.
Latent Diffusion Models (LDMs), such as Stable Diffusion, which run the diffusion process in a low-dimensional latent space to save computational resources.
Score-based Generative Modeling and Stochastic Differential Equations (SDEs) to understand the continuous-time generalization of diffusion models.
98.8K views2.5Klikes29:29@Explaining-AIOriginal Release: 2023-11-17

Denoising Diffusion Probabilistic Models (DDPM) are generative models that learn to reverse a forward diffusion process, which progressively adds Gaussian noise to images until they become pure random noise. The model learns a neural network to predict either the original image or the noise added at each step, enabling image generation by iteratively denoising random noise through the learned reverse process. The training objective is derived from maximizing the likelihood of the data, which simplifies to minimizing the squared error between predicted and actual noise or image values.