At matched compute on an RTX 5070 Ti, D-FINE is the more accurate family by a wide margin: it averages 6.4 mAP points higher and wins all 3 nearest-parameter pairs. RT-DETR answers on speed, running 33.8% faster on average in PyTorch. The choice is accuracy against throughput.
D-FINE fields 5 measured variants; RT-DETR fields 7. Both run the same COCO val2017 protocol at 640 px on an RTX 5070 Ti, so the two ladders align by parameter count.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| D-FINE-N | D-FINE | 3.8 | 4575.0 | 32.5 |
| D-FINE-S | D-FINE | 10.3 | 5339.0 | 34.2 |
| D-FINE-M | D-FINE | 19.6 | 5782.0 | 28.8 |
| RT-DETR-R18 | RT-DETR | 20.2 | 4979.0 | 47.2 |
| D-FINE-L | D-FINE | 31.2 | 5996.0 | 21.9 |
| RT-DETR-R34 | RT-DETR | 31.4 | 5223.0 | 40.2 |
| RT-DETR-L | RT-DETR | 32.9 | 5577.0 | 27.7 |
| RT-DETR-R50m | RT-DETR | 36.6 | 5382.0 | 35.3 |
| RT-DETR-R50 | RT-DETR | 42.9 | 5588.0 | 29.8 |
| D-FINE-X | D-FINE | 62.6 | 6143.0 | 18.1 |
| 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. Pairing by nearest parameter count, D-FINE wins all 3 pairs and averages 6.4 mAP points higher. At the top end D-FINE-X reaches 61.4 mAP against RT-DETR-R101 at 56.8. This is the largest accuracy gap in this set of family comparisons.
Speed
RT-DETR is the faster family: 33.8% faster on average in PyTorch FP32. Its smallest measured variant, RT-DETR-R18, runs 47.2 FPS against D-FINE-N at 32.5.
- D-FINE
- Apache-2.0 (permissive)
- RT-DETR
- Apache-2.0 (permissive)
- Commercial use
- Both families ship under Apache-2.0
Which family to pick
Pick D-FINE when accuracy leads: it is more accurate at every matched pair and reaches 61.4 mAP at the top. Pick RT-DETR when throughput is the constraint, since it runs faster at matched compute and fields more variants, 7 to 5. Both ship under Apache-2.0. See the per-variant specs on the D-FINE and RT-DETR model pages.
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.