All guides
Concept· 3 min

How does diffusion work?

Diffusion models learn to reverse noise: train by destroying an image, then generate by denoising from pure noise.

Key idea

Forward process adds Gaussian noise step by step. The model learns the reverse: predict the noise, subtract it, repeat.

Example

Stable Diffusion starts from random noise conditioned on your prompt and denoises for ~20-50 steps into a coherent image.

What to remember

  • Powers image, video and even audio generation
  • Latent diffusion works in a compressed space for speed
  • Guidance (CFG) balances prompt fidelity vs diversity