The primary objective of this study is to evaluate selected machine learning architectures for spatial localization of leaks in water distribution networks using pressure telemetry data in a multi-topology benchmark spanning 35 networks. Since all simulated scenarios contain at least one leak, scenario-level leak presence detection is trivial in this setting. Accordingly, we do not report leak/no-leak detection metrics and focus on localization selectivity. The comparative analysis focuses on node-level localization performance, spatial concentration of false alarms expressed as topological hop-distance, and inference time. The graph-based architectures exhibited a pronounced over-prediction tendency: although they achieved very high node-level recall, they also generated a large number of false-positive node classifications, limiting their practical applicability in the investigated configuration. Among the evaluated architectures, the LSTM model provided the most favorable trade-off between localization sensitivity, false-positive rate, spatial precision, and computational efficiency. It achieved a node-level sensitivity of 93.44%, a node-level false-positive rate of 21.96%, a mean false-positive hop-distance of 1.26, and an inference time of approximately 2 seconds. The results indicate that LSTM-based architectures are promising candidates for further validation toward decision-support systems for water distribution network monitoring, while graph-based models require additional probability calibration, threshold optimization, and sparsity-oriented regularization. These results should be interpreted as a comparative simulation benchmark under full-node pressure observability and per-network training; validation on real, sparsely instrumented SCADA telemetry remains future work.