Abstract: Low-resolution flow field data contains limited information, which fails to fully capture the detailed evolutionary processes of the flow field. Especially for the random turbulent features and small-scale vortex details in turbulence, they are even more challenging to obtain, thereby restricting the in-depth investigation of flow field evolution mechanisms. In order to address this limitation and reconstruct high-resolution data from low-resolution flow fields, this paper proposes a generative diffusion model called FlowDiffusionNet for flow field super-resolution reconstruction. The model takes the low-resolution flow field data input as the constraint condition, and utilizes a denoising fraction matching method to reproduce high-resolution flow field data. FlowDiffusionNet's structural design takes into consideration both the low-frequency information and high-frequency spatial features of flow field data, employing a diffusion-based modeling technique to reconstruct the residuals for high-resolution data. The proposed model's architecture is amenable to transfer learning, allowing its application to degraded flow fields at different levels. The performance of FlowDiffusionNet is evaluated on various classical flow field datasets and compared against other methods such as bicubic interpolation, super-resolution generative adversarial network (SRGAN), and super-resolution convolutional neural network (SRCNN). The results demonstrate that the proposed method achieves the best reconstruction performance on various flow fields, especially for flow field data with small-scale vortex structures down sampled by a factor of 4, where the objective evaluation index structural similarity index measure (SSIM) reaches 0.999.