Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
We analyze the Karras et al. (2022) first-order EDM diffusion sampler by decomposing its 2-Wasserstein error into universal, data-independent local discretization error (quadratic in step size) and data-dependent error amplification. We derive an explicit contraction criterion at high noise levels and measure realized amplification at low noise levels, showing it can be significantly less than the worst-case Lipschitz constant. This analysis leads to a global error bound of $O(e^{Λ_K}/K)$ and is validated by experiments on synthetic and real-world diffusion models.
Diffusion models can achieve arbitrarily low error amplification at low noise levels, decoupling sampler performance from worst-case Lipschitz bounds.
We analyze the first-order deterministic diffusion sampler of Karras et al. (2022), termed EDM, in 2-Wasserstein distance by separating two sources of error: local discretization error and its amplification by subsequent learned steps. We prove that local error admits a universal bound: for any data distribution with finite second moment, the one-step discretization error is quadratic in the step size, with an explicit constant that does not depend on the data distribution. Error propagation, in contrast, depends on the learned network. At high noise levels, we exploit the network parametrization of EDM to derive an explicit contraction criterion. At low noise levels, we measure propagation through the amplification realized on the distributions transported by the sampler; this realized amplification can be arbitrarily smaller than the worst-case Lipschitz constant. This analysis yields an $O(e^{Λ_K}/K)$ global discretization error for $K$ sampling steps, where $Λ_K$ is the low-noise log-amplification. Experiments on a one-dimensional Gaussian mixture show how measured amplification accounts for slower error decay on finite sampling grids. Diagnostics on a pretrained CIFAR-10 model illustrate related stability mechanisms without certifying the global assumptions.