Forward diffusion adds noise step by step; reverse denoises back to the original signal.
Diffusion models define a Markov chain that gradually adds Gaussian noise: q(xₜ|xₜ₋₁) = N(xₜ; √(1-β)xₜ₋₁, βI). The noise schedule ᾱₜ = ∏αₛ controls signal-to-noise ratio. Reverse denoising pθ(xₜ₋₁|xₜ) is learned to recover the original from pure noise.