On an RTX 5070 Ti in PyTorch, D-FINE averages 6.2 mAP points higher than RT-DETRv2 at matched parameter counts, winning all 3 nearest-params pairings. RT-DETRv2 is 39.7% faster on average. D-FINE leads accuracy from smallest to largest; RT-DETRv2 wins on throughput.
D-FINE fields 5 measured variants and RT-DETRv2 fields 5, all on the same COCO val2017 protocol on an RTX 5070 Ti. Three pair closely by parameter count, giving 3 matched comparisons.
| 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-DETRv2-R18 | RT-DETRv2 | 20.2 | 5075.0 | 51.1 |
| D-FINE-L | D-FINE | 31.2 | 5996.0 | 21.9 |
| RT-DETRv2-R34 | RT-DETRv2 | 31.4 | 5318.0 | 40.4 |
| RT-DETRv2-R50m | RT-DETRv2 | 36.6 | 5478.0 | 39.9 |
| RT-DETRv2-R50 | RT-DETRv2 | 42.9 | 5571.0 | 32.1 |
| D-FINE-X | D-FINE | 62.6 | 6143.0 | 18.1 |
| RT-DETRv2-R101 | RT-DETRv2 | 76.6 | 5677.0 | 25.7 |
Accuracy at matched compute
mAP is shown in percent. Each D-FINE variant is paired with the nearest RT-DETRv2 variant by parameter count. Across the 3 pairings D-FINE averages 6.2 mAP points higher and wins all three. D-FINE-X reaches 61.4 mAP at the top; RT-DETRv2-R101 reaches 56.8. At the small end D-FINE-N is a 3.78M model at 45.8 mAP, lighter than RT-DETRv2-R18 at 20.18M and 50.7 mAP.
Speed
Averaged across the matched pairs, RT-DETRv2 is 39.7% faster than D-FINE in PyTorch. Its lead holds across the range, from the D-FINE-M pairing to the D-FINE-X pairing.
- D-FINE license
- Apache-2.0
- RT-DETRv2 license
- Apache-2.0
- Both families
- permissive, cleared for commercial use
Which family to pick
Pick D-FINE for accuracy per parameter: it wins every matched pairing and reaches 61.4 mAP. Pick RT-DETRv2 for throughput, at 39.7% faster on average here. 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.