<p>In multimode fiber (MMF)-based imaging, complex speckle patterns form at the output end because of multimode interference. The input–output relationship of an MMF can be described by a transmission matrix, which strongly depends on the experimental environment and presents implementation constraints. Although deep-learning-based reconstruction methods demonstrate high reconstruction performance, they require large amounts of experimentally measured paired data, and their performance strongly depends on the characteristics of training data. In this study, we treated the MMF input–output relationship as an effective intensity-domain mapping and investigated the feasibility of approximating this mapping while preserving the statistical agreement with experimental speckle images. Using experimentally measured pairs of input images and output speckle images, we trained a U-Net-based input-to-speckle (ItoS) forward model. Quantitative evaluation revealed reasonable statistical agreement between generated and experimental speckle images in terms of intensity distribution, speckle contrast statistics, and spatial correlation. Furthermore, training the speckle-to-image inverse model using the speckle images generated by the trained ItoS forward model improved the reconstruction performance under the evaluated conditions, indicating that the generated speckle images can provide complementary support for inverse-model training. These results experimentally demonstrated that a deep-learning model can achieve a data-driven approximation of forward mapping with useful statistical agreement.</p>