Gaussian noise is random variation drawn from a bell-shaped probability distribution. Values usually cluster around 0, while increasingly extreme values become less likely. section 12.9
In diffusion models, Gaussian noise supports a learnable noising-and-denoising process:
- Sample random noise from a standard normal distribution. section 12.9
- Mix the noise with a clean image using weights controlled by a noise schedule. section 12.9
- Train the model to recognize the added pattern so it can remove noise during reverse diffusion. section 12.9
Noising an image step by step
For each timestep, sample a random noise array, calculate the square-root weights from alpha-bar, and combine the clean image with the noise. Repeating this process produces progressively noisier images while preserving the exact noise pattern for training. section 12.9
Figure 12.16 A view of probability of selection in a normal distribution
Use the normal-distribution figure to see why values near zero are selected more often than extreme values. section 12.9
