RF-DETR

Designed for fine-tuning on custom datasets, RF-DETR pairs a DINOv2 vision transformer backbone with a DETR style detection head and uses neural architecture search to trade accuracy against latency across six sizes from Nano to 2XLarge. The 2XLarge model reaches 60.1 COCO AP, while the Nano variant runs at about 2.3 ms per image on an NVIDIA T4 with TensorRT and FP16 at batch size 1. Beyond boxes, the family adds six instance segmentation variants and a preview keypoint model behind one consistent API. Training and inference come through the rfdetr Python package, installed with pip on Python 3.10 or newer, with export paths to ONNX and TensorRT for edge devices such as NVIDIA Jetson. The package and the Nano through Large detection checkpoints are Apache 2.0, while the XL and 2XLarge detection models arrive in an rfdetr_plus extension under Roboflow's PML 1.0 license, and every segmentation size stays Apache 2.0. The work was accepted to ICLR 2026, the repository has about 8,800 GitHub stars, and it is a common starting point for teams that want DETR accuracy at real-time speeds.