<p>The application of machine learning in fluid dynamics has become increasingly prevalent for accelerating computations in solving forward and inverse problems governed by partial differential equations, particularly flow field reconstruction. However, existing end-to-end methods often depend heavily on specific low-fidelity patterns or sparsity rates during training, limiting their effectiveness in real-world scenarios where inputs may deviate from the training distribution or contain unanticipated noise. Diffusion models present a promising alternative by learning to transform various low-fidelity distributions into high-fidelity ones, offering greater flexibility than direct mapping approaches that are typically constrained to single problem types. In this work, we introduce the physics-informed residual diffusion model, which generalizes across diverse inputs, including evenly down-sampled sensor data, sparse sensor data with Gaussian noise, and randomly sampled sparse sensor data. By incorporating partial differential equation constraints into the objective function, our approach significantly improves the accuracy of reconstructed high-fidelity flow fields while ensuring adherence to underlying physical laws. Experimental results demonstrate that our model successfully generates high-quality outcomes for two-dimensional Navier-Stokes equation flow and Kolmogorov flow under varied low-fidelity input conditions, without the need for retraining.</p>