Does Semantic Noise Initialization Transfer from Images to Videos? A Paired Diagnostic Study

TL;DR

This study investigates whether semantic noise initialization improves text-to-video diffusion models, finding limited transferability and emphasizing the need for careful evaluation and diagnostics in this domain.

Contribution

It provides the first benchmark comparing semantic noise initialization to Gaussian noise in T2V models, highlighting its limited benefits and proposing evaluation standards.

Findings

Small positive trend in temporal dimensions, but not statistically significant

Overall performance remains comparable to baseline

Noise perturbation analysis reveals weak or unstable signals

Abstract

Semantic noise initialization has been reported to improve robustness and controllability in image diffusion models. Whether these gains transfer to text-to-video (T2V) generation remains unclear, since temporal coupling can introduce extra degrees of freedom and instability. We benchmark semantic noise initialization against standard Gaussian noise using a frozen VideoCrafter-style T2V diffusion backbone and VBench on 100 prompts. Using prompt-level paired tests with bootstrap confidence intervals and a sign-flip permutation test, we observe a small positive trend on temporal-related dimensions; however, the 95 percent confidence interval includes zero (p ~ 0.17) and the overall score remains on par with the baseline. To understand this outcome, we analyze the induced perturbations in noise space and find patterns consistent with weak or unstable signal. We recommend prompt-level paired evaluation and noise-space diagnostics as standard practice when studying initialization schemes for T2V diffusion.