Denoising removes what a model believes is noise from an image representation. In Stable Diffusion, the prompt helps distinguish meaningful structure from unwanted variation. section 4.2.1
The useful trick is to start with pure noise. The model removes a little noise at a time, guided by the prompt, until recognizable content emerges. It is image-making by repeatedly finding shapes in a cloud of static. section 4.2.1
The workflow has a few moving parts:
- A U-Net predicts what the noise in the current image representation should be. section 5.0
- The prompt guides the prediction toward content such as a tree at night, rather than merely producing a cleaned-up image. section 4.2.1
- Repeated sample steps gradually move the result toward the requested image. section 4.2.1
- A sampler controls how much noise disappears at each step, while a scheduler determines the step size. section 6.0
From static to structure
Give Stable Diffusion random noise and the prompt “a tree at night.” Early passes find a rough silhouette; later passes develop features such as branches and other details. section 4.2.1








