This paper proposes HazeMatching, a computational dehazing technique, to address the problem of blurry image data from inexpensive and accessible wide-field microscopes. HazeMatching uses an iterative approach, modifying the conditional flow matching framework to incorporate blurry observations into the conditional velocity field, aiming to balance data accuracy (MSE, PSNR) and perceptual realism (LPIPS, FID). We evaluate our method against seven existing methods on five datasets, including synthetic and real data, and demonstrate that it effectively balances accuracy and realism while generating well-calibrated predictions. Its applicability to real-world microscopy data without the need for explicit degradation operators is a key advantage, and the data and code used will be made publicly available.
