Validate the theoretical error-propagation conditions

Measure the direct Jacobian gains and cross-step cosine similarities associated with persistent spatial weight perturbations in diffusion sampling, to validate the sufficient conditions used to explain the greater vulnerability of the early denoising stage.

Background

The paper derives a first-order error recursion for fixed spatial weight noise and uses it to explain why perturbations introduced during the early, high-noise denoising stage can propagate more strongly than perturbations introduced during final refinement. This explanation relies on propagation gains and nonnegative cross-step alignment assumptions within activation windows.

The authors evaluate the resulting behavior using stage-wise FID, but do not directly measure the Jacobian-based propagation gains or the cosine similarities between propagated errors. Such measurements would provide a more direct empirical test of the theoretical mechanism underlying ASSERT’s temporal stochasticity schedule.

References

We evaluate the resulting prediction through stage-wise FID, while direct Jacobian-gain and cross-step-cosine measurements are left for future work.

— ASSERT: Adaptive Stochastic Sampling for Robust Diffusion Models on Analog Compute-in-Memory Hardware  (2609.00955 - Feng et al., 1 Sep 2026) in Section 3.2, “Persistent-Noise Error Analysis,” subsection “Sufficient condition for early-stage vulnerability”