Small-object detection plays a critical role in real-time aerial, Unmanned Aerial Vehicle (UAV), and satellite vision systems, as well as many other domains, where hardware constraints and deployment environments demand high accuracy, low latency, and energy efficiency. Although numerous lightweight object detectors have been proposed, there exists a lack of rigorous, cross-platform benchmarks that evaluate these models under consistent conditions for small-object-focused applications. This paper presents a comprehensive comparative evaluation of eight recent lightweight You Look Only Once (YOLO)-based models, including YOLOX-Nano and YOLOv6–v12 variants. Models are trained using a unified pipeline and evaluated on three representative datasets: VisDrone2019, DOTA-v1.5, and xView. Performance is assessed on both cloud GPUs (NVIDIA A100) and an edge AI platform (Jetson Orin Nano), using PyTorch and TensorRT (FP16) runtimes. The evaluation metrics include mAP@0.5, mAP@0.5:0.95, F1-score, latency, FPS, memory footprint, and energy consumption (via VDD_IN telemetry). Experimental results demonstrate that YOLOv8-Nano and YOLOv9-Tiny achieve leading accuracy across datasets, while YOLOv6-Nano and YOLOv10-Nano offer favorable trade-offs in speed and memory. TensorRT optimization yields 3×–5× speedup over PyTorch. Energy draw ranges from 7.6 W to 10.4 W, depending on the model and configuration. Pareto frontiers reveal that no single model dominates across all objectives. The study highlights the importance of multi-objective benchmarking in edge AI and provides practical guidance for selecting detectors based on deployment-specific constraints. The findings support theory-driven and application-oriented advancement of efficient detection architectures for real-time, resource-aware computer vision systems.