<![CDATA[YOLOv12 is one of the latest YOLO versions currently. Several studies have proven that YOLOv12 has better performance compared to previous versions. YOLOv12 itself has five model variants based on its architectural complexity, namely nano, small, medium, large and extra larges. This study tests the performance of YOLOv12 model variants (n, s, m, l, x) for traditional Indonesian culinary detection using a domain-specific object detection dataset. The dataset contains 718 images with 720 bounding-box instances annotated across 20 culinary classes, divided into 418/150/150 images for training/validation/testing. Data processing was performed in Roboflow with automatic orientation and stretching resizing to 640×640, while the training split was enriched using augmentation (horizontal and vertical flips) to increase sample diversity. All YOLOv12 variants were trained with the same configuration and environment, for 50 epochs using the Ultralytics framework with default hyperparameters on an NVIDIA A100-SXM4 80GB GPU. On the validation set, all variants achieved high detection accuracy (mAP@0.5 = 0.985–0.991), while differences emerged under a more stringent localization criterion (mAP@0.5:0.95). The best overall localization performance was achieved by YOLOv12-L (mAP@0.5:0.95 = 0.874), while YOLOv12-N provided the fastest inference (0.8 ms/image) with competitive accuracy (mAP@0.5:0.95 = 0.822). These findings provide preliminary guidance for selecting YOLOv12 variants based on the trade-off between accuracy and speed.]]>
