This video provides a comprehensive explanation of score-based diffusion models, presenting them as a continuous-time alternative to the more common discrete DDPM framework. It details how the forward noising process, which corrupts data into Gaussian noise, can be formally described by a Stochastic Differential Equation (SDE). The core of the generation process relies on a known mathematical theorem that guarantees the existence of a corresponding reverse-time SDE. This reverse SDE's drift term is dependent on the "score function"βthe gradient of the log-probability of the data distribution at each point in time. The video explains how a neural network is trained to estimate this score, which conveniently simplifies to a noise prediction task similar to DDPMs. Finally, it highlights the advantages of this SDE formulation, such as its flexibility in choosing different noising processes and numerical solvers, which provides a more general and extensible foundation for generative modeling.
A Deep Dive into Score-Based Diffusion Models
0:00:27 Two Views of Diffusion: The video explains that there are two primary mathematical frameworks for diffusion models: the discrete Denoising Diffusion Probabilistic Models (DDPM) and a continuous-time version called score-based diffusion models, which use Stochastic Differential Equations (SDEs).
0:02:35 The Forward Process as an SDE: Instead of discrete steps, the process of adding noise to an image is modeled as a continuous SDE. This equation consists of a deterministic "drift" term that pulls the data towards the origin and a random "diffusion" term that adds noise, gradually transforming any data distribution into a simple Gaussian one.
0:09:19 Reversing the Process: A crucial mathematical theorem states that the forward SDE has a corresponding reverse-time SDE. This allows the process to be run backward, starting from pure Gaussian noise and gradually transforming it back into a sample from the original data distribution (e.g., a realistic image).
0:10:11 The Role of the Score Function: The reverse SDE requires knowledge of the "score function" at every time t. This function is a vector that points in the direction of the highest probability density, essentially guiding the noisy data back towards valid image manifolds.
0:12:24 Training a Network to Learn the Score: Since the true score function is unknown, a neural network (like a U-Net) is trained to estimate it. The network takes a noisy image and the corresponding time step t as input and outputs an estimate of the score.
0:14:10 The Practical Training Objective: While the conceptual goal is to learn the score, the training objective simplifies to a much easier task: predicting the noise (epsilon) that was added to the original clean image. This makes the training process nearly identical to that of a DDPM.
0:15:23 Generating Images with a Numerical Solver: To generate a new image, the trained score network is used within the reverse SDE. This continuous equation is solved iteratively using a numerical method (like the Euler-Maruyama solver) to go from a random noise sample at time T to a clean image at time 0.
0:17:40 Key Advantages of the SDE Framework: The score-based SDE formulation is more general and flexible than the discrete DDPM approach. It allows for the use of many different noising processes and a variety of advanced numerical solvers, which can lead to faster or more accurate image generation.