This study investigates the application of YOLO (You Only Look Once) models for detecting exomoons in astronomical images. Multiple experiments were conducted using different simulations and labeling strategies. Notably, using labels based on the signal-to-noise ratio while imitating realistic fluxes demonstrated intriguing potential in its application to real data. The training experiments unveiled nuanced challenges, with experiments showcasing robust performance on simulated data but encountering complexities with under-subtraction when transitioning to real data. Despite being trained with fewer objects, subsequent experiments aimed at refining detection on real data present notable advancements. The analysis sheds light on under-subtraction issues, emphasizing the potential influence of multiple classes on object characteristics and bounding box predictions.