Abstract: Recently, large-scale text-to-image diffusion models have shown significant potential in text-guided video editing. However, due to the lack of precise temporal modeling, the generated videos often suffer from se-vere temporal inconsistencies. To address this challenge, we propose a novel text-guided video editing method based on a distribution-preserving inter-frame noise propagation mechanism. Specifically, we first employ DDIM inversion to map the source video into latent prior noise. Then, within the same timestep of the denoising phase, we perform a weighted aggregation of the predicted noise samples from the current and adjacent frames, explicitly enabling effective inter-frame information propagation. Furthermore, a strict variance-preserving constraint is introduced during the aggregation to ensure the propagated noise statistically conforms to the original Gaussian distribution, thereby avoiding the disruption of the baseline model's denoising stability. Finally, this training-free mechanism is seamlessly integrated into existing baselines to generate high-quality edited videos guided by target prompts. Extensive experiments on da-tasets such as LOVEU-TGVE demonstrate that, without introducing additional computational overhead, our method improves the temporal consistency, textual alignment, and human preference scores of baselines by 2.61, 1.74, and 0.28, respectively. Moreover, it exhibits superior performance in multi-object editing and generalizes well to large-scale text-to-video models.