On an RTX 5070 Ti in PyTorch, RF-DETR averages 1.6 mAP points higher than RT-DETR at matched parameter counts and wins 3 of the 4 nearest-params pairings; the one it loses is the smallest. RT-DETR runs about 3% faster on average and fields more variants. The choice is accuracy per parameter against throughput and range.
RF-DETR fields 4 measured variants and RT-DETR fields 7, all evaluated on the same COCO val2017 protocol on an RTX 5070 Ti. Pairing each RF-DETR variant with the nearest RT-DETR variant by parameter count gives 4 matched comparisons.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| RT-DETR-R18 | RT-DETR | 20.2 | 4979.0 | 47.2 |
| RF-DETR-N | RF-DETR | 30.5 | 5135.0 | 39.5 |
| RT-DETR-R34 | RT-DETR | 31.4 | 5223.0 | 40.2 |
| RF-DETR-S | RF-DETR | 32.1 | 5512.0 | 35.8 |
| RT-DETR-L | RT-DETR | 32.9 | 5577.0 | 27.7 |
| RF-DETR-M | RF-DETR | 33.7 | 5732.0 | 30.3 |
| RF-DETR-L | RF-DETR | 33.9 | 5855.0 | 25.1 |
| RT-DETR-R50m | RT-DETR | 36.6 | 5382.0 | 35.3 |
| RT-DETR-R50 | RT-DETR | 42.9 | 5588.0 | 29.8 |
| RT-DETR-X | RT-DETR | 67.4 | 5794.0 | 21.8 |
| RT-DETR-R101 | RT-DETR | 76.6 | 5677.0 | 23.6 |
Accuracy at matched compute
mAP is shown in percent. The comparison pairs each RF-DETR variant with the nearest RT-DETR variant by parameter count. Across the 4 pairings RF-DETR averages 1.6 mAP points higher and wins 3; RT-DETR wins 1. At the top end RF-DETR-L reaches 58.6 mAP against RT-DETR-R101 at 56.8. At the small end RF-DETR-N (30.47M, 51.4 mAP) faces the lighter RT-DETR-R18 (20.18M, 49.8 mAP).
Speed
Speed runs the other way, but narrowly. Averaged over the matched pairs, RT-DETR is about 3% faster than RF-DETR in PyTorch. The gap swings by pairing: RF-DETR-M leads its match by 9.4%, while RF-DETR-L trails by the same margin.
- RF-DETR license
- Apache-2.0
- RT-DETR license
- Apache-2.0
- Both families
- permissive, cleared for commercial use
Which family to pick
Pick RF-DETR when accuracy per parameter matters: it leads on average and tops out at 58.6 mAP with its largest variant. Pick RT-DETR for throughput or a wider size ladder, since it fields 7 variants to RF-DETR's 4. Both ship under Apache-2.0, so licensing does not force the call.
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.