Blog Post··6 min read
Score Matching: The Math Behind Diffusion
Diffusion models learn to reverse a noise process. The key insight is that you don't need to know the data distribution — you only need to learn its score function, the gradient of the log-density.