Evaluating model quantization in a GenAI-enhanced weed detection pipeline

Deep learning–based weed control systems often struggle with limited training data diversity and constrained computational resources, restricting their effectiveness in real-world deployment. To address these limitations, we introduce a Stable Diffusion–based inpainting framework that progressively augments training datasets in 25% increments, up to 200%, enriching both data volume and variability. We systematically evaluate three state-of-the-art object detection architectures, such as large, small, and nano variants of YOLO11 and YOLOv12, along with large RT-DETR models, under three precision settings (FP32, FP16, INT8) using mAP50 and mAP50-95 evaluation metrics. Experiments on NVIDIA Jetson Orin Nano, NVIDIA Jetson AGX Orin, and spo-comm rugged computing unit reveal that quantization consistently reduces latency and memory footprint, with INT8 compression producing the most compact and fastest models. While INT8 often induces accuracy degradation, we show that this loss is significantly minimized by targeted synthetic augmentation. Notably, small YOLO variants trained with augmented data match, and in some cases surpass, the detection performance of their baseline large counterparts, without added model size or inference cost. Furthermore, utilizing the INT8-quantized Stable Diffusion for data generation preserves augmentation benefits on the downstream models while minimizing generation overhead. In combination, these contributions establish a novel training and deployment strategy for embedded AI in the context of weed detection, demonstrating that small YOLO models, INT8 quantization, and targeted synthetic augmentation can jointly deliver higher efficiency without sacrificing accuracy.