IC-1579ADM's noise prediction network exhibits exposure bias: during iterative sampling the l2-norm of its ε prediction is systematically larger than during training, and the sampling distribution variance exceeds the training variance with error accumulating toward the end of the chain
Mang Ning, Mingxiao Li, Jianlin Su, Albert Ali Salah, Itir Onal Ertugrul
The paper analytically derives that the prediction error x̂t_θ − x0 causes the sampling distribution q_θ(x̂t|x_{t+1}) to have larger variance than the training distribution q(xt|x0) by the term (√α_t β_{t+1}/(1−α_{t+1}) · e_{t+1})². This is empirically verified on the pre-trained ADM model on CIFAR-10 using 20-step sampling with 50k samples per timestep. The l2-norm of εθ during sampling is consistently larger than during training at every timestep, and the multi-step variance error grows monotonically toward t=1 (the end of sampling), confirming that prediction error accumulates over the sampling chain.
Evidence
observational
Caveat
The analysis assumes the prediction error can be modelled as a Gaussian with mean x0; the paper notes this assumption does not affect the variance gap conclusion but the specific magnitude of the error term depends on this approximation.