DL0127 DDPM Training Objective

A denoising diffusion probabilistic model is trained by maximizing a variational bound on the data log-likelihood, and because both the forward corruption and the reverse posterior are Gaussian, that bound collapses into a sum of KL divergences whose only learnable content is a mean-matching term at each noise level. Reparameterizing that mean through the closed-form forward marginal turns every term into a weighted squared error between the injected noise and a network prediction, and Ho et al. then drop the per-step weight to get the simple objective . Predicting the noise is not a different model from predicting the clean image: -prediction, -prediction, and direct mean prediction are affine reparameterizations of one another given and , so they define the same optimum but different implicit loss weightings and different numerical conditioning. The practical reasons to regress noise are that the target is always unit-variance no matter how noisy the input is, that the residual structure lets the network pass the low-frequency content through instead of re-synthesizing it, and that discarding under this parameterization down-weights the easy low-noise steps by roughly an order of magnitude, which empirically buys much better FID. It also makes the network a rescaled score estimator, which is what connects DDPM to score matching and to every ODE and SDE sampler built on top of it.

(1) The Objective Is One Scalar MSE: sample , sample and , form in closed form, and regress . No per-step model, no adversarial term, no sequential rollout during training.
(2) It Comes From The ELBO, Not From Heuristics: each is a Gaussian KL whose learnable part is a mean difference, and substituting the forward marginal converts it into .
(3) Dropping Is A Reweighting Choice: is largest at small , so setting all weights to 1 shifts capacity toward the high-noise steps that decide global structure.
(4) The Target Is Scale-Stable: is standard normal for every , so a single output head with fixed normalization works across the whole schedule, whereas the difficulty of predicting varies enormously with .
(5) Residual Prediction Is Easier For A U-Net: at low noise, most of is already the answer, so predicting the small perturbation avoids forcing the network to reconstruct an image it was handed.
(6) Conditioning Flips At The Two Ends: converting to amplifies error by , which is tiny at low noise and roughly 150x at , the reason v-prediction exists.

Figure 1: One training step. The closed-form forward marginal means an arbitrary timestep can be sampled directly, so training never simulates the chain; the loss is a single unit-variance regression shared by all noise levels.

The reweighting argument is the one that most candidates miss. Under the ELBO, the term at carries roughly 50x the weight of the term at for the standard linear schedule, and those low- terms correspond to removing almost imperceptible noise, a task that contributes little to perceptual quality but a lot to likelihood. Setting every weight to 1 in -space is therefore a deliberate trade of likelihood for sample quality, and it is why DDPM reports strong FID with mediocre bits-per-dimension. The conditioning argument cuts the other way at high noise: since is recovered by dividing by , a small noise error becomes a large image error near , so pure -prediction is a poor target for few-step or distilled samplers that must produce a usable from the very first step. Finally, the identity shows the trained network is a scaled score function, which is exactly what a probability-flow ODE or an annealed Langevin sampler needs, so the same checkpoint serves DDPM, DDIM, and higher-order solvers.

Mathematical Formulation:

Where:

• is a clean training sample, its corrupted version at step , and the injected noise that serves as the regression target.
• is sampled uniformly, is the schedule, , and gives the closed-form marginal that makes single-step training possible.
• is the U-Net with parameters , conditioned on through a sinusoidal timestep embedding.
• is the ELBO weight and the reverse-process variance (commonly ); is with replaced by 1.
• is the signal-to-noise ratio, which decreases monotonically in and sets the error amplification from noise space into image space.
• is the score, so a trained doubles as a score model for ODE and SDE samplers.

Figure 2: The two quantitative reasons the parameterization matters. Left: dropping removes a roughly 50x preference for the easiest low-noise steps. Right: the same trained error costs almost nothing in image space at low noise but is amplified by near , which is where v-prediction stays bounded.