Score model in diffusion models

In this post, I will introduce the score model in diffusion models. I recommend reading my previous posts before proceedings.

Reverse SDE in diffusion models

Score function of a stochastic process

Reverse stochastic differential equation (SDE)

In diffusion models, data samples are generated by integrating reverse SDE:

where \nabla_{x_t} \log p_t(x_t) is the score function of x_t .

Score model in diffusion model

Since the score function in (Reverse SDE) is often intractable, it is approximated by a neural network denoted as s_\theta (\cdot) , which is parametrized with \theta . This neural network s_\theta(\cdot) is called the score model.