In this paper, we propose HazeMatching, a computational dehazing technique to address the problem of blurry image data from inexpensive and accessible wide-field microscopes. HazeMatching guides the generation process by incorporating blurry observations into conditional velocity fields using a conditional flow matching framework, aiming to strike a balance between data fidelity and realism. We demonstrate that it consistently achieves a balance between fidelity and realism by evaluating it against seven baseline models on five datasets containing synthetic and real data. Furthermore, we verify through calibration analysis that HazeMatching generates well-calibrated predictions and can be easily applied to real microscope data without explicit degradation operators. All data and code are publicly available.
