High-frequency temporal reconstruction of temperature fields aims to recover sub-daily temperature evolution from low-frequency states,which is essential for representing rapidly changing thermal variations. However, most existing methods formulate this task as a direct estimation of intermediate temporal states without explicitly modeling the latent transition process connecting adjacent states, which limits their ability to capture underlying dynamics and may lead to temporally inconsistent predictions. To address this limitation, we propose ITDiff, a motion-aware diffusion framework for high-frequency temporal reconstruction of temperature fields, in which Implicit Temperature Flow Fields (ITFFs) are introduced to represent the latent transition process between boundary and intermediate states. First, a Flow Teacher estimates ITFFs from the boundary and target states; these ITFFs are then used to guide a Synthesis Model in reconstructing temporally consistent intermediate states. Building on this teacher-guided representation, a Conditional Diffusion Model is subsequently trained to infer plausible ITFFs from boundary conditions alone, thereby enabling the recovery of fine-grained temporal variations that are often oversmoothed by conventional approaches. Extensive experiments demonstrate that ITDiff consistently outperforms representative baselines, improving RMSE-based reconstruction accuracy by up to 48 % while better preserving temporal structures across multiple scales.