Diffusion models are a kabbalistic miracle of math. At the core, they’re just incredibly advanced denoising systems, formally known as Denoising Diffusion Probabilistic Models (DDPMs), e.g. Stable Diffusion and DALLE-2.

During training, the model is shown hundreds of millions of images paired with text descriptions. To teach it how to "clean up" noisy images, we intentionally add random noise to each training image. The model’s job is to learn how to reverse it using the text prompt as a guide for where and how to remove the noise.

When you generate an image, the model performs this process in reverse. It starts with a latent space of pure random noise and gradually subtracts more and more noise with each diffusion step. It's synthesizing an image from scratch by removing all of the noise until the image remains, organizing the chaos into whatever you asked it to generate.