In fields such as Industry 4.0 and smart cities, digital twin technology is being widely applied. Constructing virtual models that correspond directly to physical entities, enables real-time monitoring, predictive maintenance, and optimized management of devices, products, and systems. However, efficiently generating accurate network traffic packets for digital twins remains a highly challenging task. This task involves temporal dependencies, and under the influence of complex network behaviors and diverse scenarios, the traffic generation process is a network traffic packet generation algorithm—FlowDiff—and a Temporal Diffusion Generation Model (TDGM). The FlowDiff algorithm treats the generation of traffic packets as a reverse diffusion process, where, based on the temporal characteristics of traffic and external scene conditions, noise is progressively removed to generate traffic packets that match real-world network behavior. The TDGM model is designed to adapt to the temporal dependencies and periodic variations in network traffic. It introduces a temporal-aware feature embedding layer, which fuses the temporal information of traffic with its features, thereby enhancing the spatiotemporal relationships between traffic characteristics. The model combines Convolutional Neural Networks (CNN) and Transformer modules to extract both local and global features of the traffic and fuse them effectively. Finally, historical traffic data and periodic features are used as conditional inputs to the diffusion model, and a cross-attention mechanism further optimizes the generation process. Experimental results show that FlowDiff performs excellently on real network traffic datasets. A comparison with real traffic data demonstrates its high accuracy and effectiveness, highlighting its potential and advantages for application in digital twin systems.