Verdict

On an RTX 5070 Ti, speed is rarely the constraint. D-FINE-X is the most accurate model here at 61.5 mAP and still runs at 73.0 FPS. If you need 120 FPS, RT-DETR-L gives 55.8 mAP at 122.0 FPS. All numbers are TensorRT FP16, batch 1, mAP@50-95 in percent.

Fastest models on the RTX 5070 Ti

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1RT-DETRv2-R3453.2132.97.5264031.0Apache-2.0
2DEIMv2-Atto25.8132.17.573200.5Apache-2.0
3RT-DETR-L55.8122.08.2064032.0Apache-2.0
4RT-DETRv2-R1850.8117.88.4964020.0Apache-2.0
5RT-DETR-R1849.8113.48.8264020.0Apache-2.0
6RT-DETRv2-R50m54.7113.08.8564036.0Apache-2.0
7RT-DETR-R50m53.9110.19.0864036.6Apache-2.0
8RT-DETR-R3452.3108.49.2264031.0Apache-2.0
9RT-DETRv2-R5055.7103.69.6564042.0Apache-2.0
10YOLOX-Nano28.7100.09.994160.9Apache-2.0
11YOLOX-Tiny35.497.710.234165.1Apache-2.0
12RT-DETR-R5055.897.410.2764042.0Apache-2.0
13RT-DETRv4-S52.892.310.8464010.0Apache-2.0
14RT-DETR-R10156.788.011.3664076.0Apache-2.0
15RT-DETRv2-R10156.887.411.4464076.0Apache-2.0
Fastest detectors on RTX 5070 Ti, TensorRT FP16, batch 1. Timings include pre and post-processing.

The fastest rows hit a floor. DEIMv2-Atto, the smallest model on the board, takes 7.57 ms per image. RT-DETRv2-R34, a much larger transformer, takes 7.52 ms. Timings are end to end, so below that floor most of the time is per-image overhead, not the network. A smaller model will not get you more frames. Batching or a leaner input pipeline will.

Most accurate models

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1D-FINE-X61.573.013.7064062.0Apache-2.0
2ECDet-X61.041.324.2264049.0Apache-2.0
3DEIMv2-X60.837.926.4064050.3Apache-2.0
4D-FINE-L60.069.314.4264031.0Apache-2.0
5ECDet-L60.041.724.0064031.0Apache-2.0
6RT-DETRv4-X60.071.014.0764062.0Apache-2.0
7DEIM-X59.570.714.1564062.0Apache-2.0
8RF-DETR-L58.647.621.01704128.0Apache-2.0
9DEIMv2-L58.538.725.8664032.2Apache-2.0
10ECDet-M58.248.920.4664018.0Apache-2.0
Most accurate detectors on RTX 5070 Ti, TensorRT FP16, batch 1.

ECDet-X and DEIMv2-X land within a point of D-FINE-X, at 41.3 and 37.9 FPS. D-FINE-X is both more accurate and faster, so on this card there is no trade-off to make at the top.

Every number on this page comes from the verified dataset: same 500-image COCO val2017 slice, conf 0.001, IoU 0.6, max 300 detections, pycocotools mAP, identical protocol across all hardware and runtimes. The full protocol is on the methodology page. To rerun this comparison with your own filters, open compare. Accuracy is measured on LibreYOLO retrained checkpoints; other weight sources can yield different values.