reverse denoising process
The reverse denoising process is the hard, learned half of a diffusion model — the act of restoration that undoes the deliberate vandalism of the forward process. You begin with a canvas of pure random noise and, step by step, scrub away a little of it, each scrub guided by a neural network that asks 'given how noisy this looks, what did the cleaner version one notch back probably look like?' After many such scrubs a sharp, coherent image emerges from what started as static. This is the process that actually generates pictures.
Formally the reverse process is also modeled as a Markov chain, p_theta(x_{t-1} | x_t) = N(x_{t-1}; mu_theta(x_t, t), Sigma_t). The network does not output the clean image directly; it typically predicts the noise epsilon_theta(x_t, t), from which the mean mu_theta of the slightly cleaner image is computed by a fixed algebraic formula. The variance Sigma_t is often fixed to the schedule's values (though it can also be learned). The deep reason this works is that when the forward steps are small enough, the true reverse conditional is itself approximately Gaussian — so a Gaussian model with a learned mean is the right functional form, and the only thing left to learn is where to center it.
Each reverse step is stochastic in the original DDPM sampler: after computing the mean, you add a small dose of fresh Gaussian noise (except at the very last step), which keeps the trajectory from collapsing and lets one starting noise map to a rich distribution of images. This is why two runs from different initial noise give different pictures, and why the same prompt can yield endless variations. The number of reverse steps trades speed against quality: more steps mean each correction is smaller and more accurate, but generation is slower — the central tension that DDIM and few-step distillation methods attack.
Why not skip the chain and have the network map noise to image in one pass, like a GAN? Because the reverse conditional is only near-Gaussian for small steps; over a giant leap it becomes a wildly multimodal distribution that a single Gaussian cannot capture, so a one-shot prediction regresses to a blurry mean. Iteration is what lets a simple Gaussian model represent an arbitrarily complex data distribution.