# Gaussian Noise Gaussian noise is random noise whose values follow a normal distribution (also called a **Gaussian distribution**). In probability theory, this means the noise is characterized by a bell-shaped probability density function defined by just two parameters: a *mean* (center) and *standard deviation* (spread). The mathematical form is as follows: *where $\mu$ is the mean and $\sigma$ is the standard deviation.* $p(z) = \frac{1}{\sigma\sqrt{2\pi}} e^{-\frac{(z-\mu)^2}{2\sigma^2}}$ ## What makes it "white" and why it matters A special case called **white Gaussian noise** is particularly useful in AI generation. "White" just means every noise value is independent; knowing one pixel's noise tells you nothing about any other. There are no clusters, streaks, or patterns baked in. This clean, patternless randomness is what makes it a useful starting point: a model trained to reverse it doesn't have to fight any hidden structure in the noise itself. ^ai-advantage ## Role in Diffusion Models In [[diffusion_models#^overview|diffusion models]], Gaussian noise is the engine of the process. A diffusion model works by either [[diffusion_models#^forward-process|Forward Process (Training)]] or [[diffusion_models#^reverse-process|Reverse Process (Generation)]]. The key insight is that small, incremental additions of Gaussian noise allow the network to learn a smooth, reversible transformation. This is why Gaussian noise is preferred: ==its mathematical properties guarantee that you can theoretically reverse the process perfectly if you learn the right denoising function==. ## Intuition Think of Gaussian noise like static on an old television. If you zoom in on pure static, you see random dots of varying brightness, centered around some average gray level. Most dots cluster near the average, with fewer at the extremes, that's the bell curve. This randomness-with-a-center is what makes Gaussian noise so useful: it's unpredictable but not chaotic. You can describe it, control it, and reverse it.