D-FINE is the more accurate family: at matched parameter counts on an RTX 5070 Ti it averages 7.5 mAP points higher than RF-DETR and wins both nearest-params pairings. It also holds the top: D-FINE-X reaches 61.4 mAP against RF-DETR-L at 58.6. The trade is speed: RF-DETR runs 35.8% faster on average in PyTorch. Pick D-FINE for accuracy and RF-DETR for throughput.
D-FINE fields 5 measured variants and RF-DETR fields 4, all on the same COCO val2017 protocol on an RTX 5070 Ti. RF-DETR's variants sit between 30.47M and 33.93M parameters, so only 2 pairings match closely by parameter count: D-FINE-M and D-FINE-L, both against RF-DETR-N.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| D-FINE-N | D-FINE | 3.8 | 45.8 | 32.5 |
| D-FINE-S | D-FINE | 10.3 | 53.4 | 34.2 |
| D-FINE-M | D-FINE | 19.6 | 57.8 | 28.8 |
| RF-DETR-N | RF-DETR | 30.5 | 51.4 | 39.5 |
| D-FINE-L | D-FINE | 31.2 | 60.0 | 21.9 |
| RF-DETR-S | RF-DETR | 32.1 | 55.1 | 35.8 |
| RF-DETR-M | RF-DETR | 33.7 | 57.3 | 30.3 |
| RF-DETR-L | RF-DETR | 33.9 | 58.6 | 25.1 |
| D-FINE-X | D-FINE | 62.6 | 61.4 | 18.1 |
Accuracy at matched compute
mAP is shown in percent. The comparison pairs each D-FINE variant with the nearest RF-DETR variant by parameter count. D-FINE-M (19.59M) beats RF-DETR-N by 6.4 points and D-FINE-L (31.24M) by 8.6, for a mean of 7.5. At the top end D-FINE-X reaches 61.4 mAP against RF-DETR-L at 58.6. At the small end D-FINE-N is a 3.78M model at 45.8 mAP, far lighter than RF-DETR's smallest at 30.47M and 51.4 mAP.
Speed
Averaged across the matched pairs, RF-DETR is 35.8% faster than D-FINE in PyTorch. RF-DETR-N runs 39.5 FPS against 28.8 for D-FINE-M and 21.9 for D-FINE-L.
- D-FINE license
- Apache-2.0
- RF-DETR license
- Apache-2.0
- Both families
- permissive, cleared for commercial use
Which family to pick
Pick D-FINE when you want the most accuracy per parameter: it leads at both matched sizes and reaches 61.4 mAP at the top. Pick RF-DETR when throughput is the constraint, since it runs 35.8% faster on average in this pairing. 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.
