2D-to-3D lifting methods have become the mainstream approach for estimating 3D human poses. While diffusion models have achieved remarkable success in single-frame image prediction, estimating 3D human poses in videos remains a significant challenge, particularly in scenes with long temporal sequences, severe occlusions, or rapid motion changes, where existing methods often struggle in terms of robustness and accuracy. In this paper, we propose DTFPose, a novel method that applies diffusion in both the temporal and frequency domains for 3D human pose estimation in videos. At its core, the Diffusion in Time and Frequency Domains (DiffTF) module fuses temporal and spectral features during the reverse diffusion process. Experimental results demonstrate that DTFPose achieves state-of-the-art performance on the Human3.6M and MPI-INF-3DHP datasets.