Skip to main content

Learn · Diffusion

Diffusion: noise to image

Image generators like Stable Diffusion learn to reverse a noising process. Forward: add Gaussian noise step by step until an image is pure static. Reverse: a model removes a little noise at each step to recover an image. Drag the slider to add noise, then hit generate to watch it denoise.

Signal kept √ᾱ
Noise added √(1-ᾱ)

What's really happening

Each pixel follows the diffusion equation xₜ = √ᾱₜ·x₀ + √(1-ᾱₜ)·ε — a shrinking amount of the original image plus growing Gaussian noise, on a cosine schedule. Training a diffusion model means teaching a network to predict the noise ε at any step, so it can subtract a little at a time and walk back from static to a clean image. Text-to-image models add a twist: a text prompt steers which image the denoising heads towards. This demo denoises a known picture; a real model hallucinates a brand-new one from pure noise.

Real diffusion runs over hundreds of steps in a compressed "latent" space, not raw pixels — but the maths of the forward process shown here is exactly the same. For the language-model side of generative AI, see the Token and Sampling demos.