On an RTX 5070 Ti in PyTorch, RF-DETR averages 3.9 mAP points higher than YOLOX at matched parameter counts and wins 3 of 4 nearest-params pairings, losing only the smallest. YOLOX is 32.2% faster on average and scales to much larger models. Pick by accuracy per parameter against throughput and range.
RF-DETR fields 4 measured variants and YOLOX fields 6, all on the same COCO val2017 protocol on an RTX 5070 Ti. RF-DETR's variants cluster near 30M parameters, so all 4 pair against YOLOX-M by count.
| Model | Family | Params (M) | mAP@50-95 | FPS |
|---|---|---|---|---|
| YOLOX-Nano | YOLOX | 0.9 | 2876.0 | 53.6 |
| YOLOX-Tiny | YOLOX | 5.1 | 3547.0 | 61.6 |
| YOLOX-S | YOLOX | 9.0 | 4427.0 | 50.0 |
| YOLOX-M | YOLOX | 25.3 | 5169.0 | 48.2 |
| RF-DETR-N | RF-DETR | 30.5 | 5135.0 | 39.5 |
| RF-DETR-S | RF-DETR | 32.1 | 5512.0 | 35.8 |
| RF-DETR-M | RF-DETR | 33.7 | 5732.0 | 30.3 |
| RF-DETR-L | RF-DETR | 33.9 | 5855.0 | 25.1 |
| YOLOX-L | YOLOX | 54.2 | 5542.0 | 42.8 |
| YOLOX-X | YOLOX | 99.1 | 5627.0 | 39.6 |
Accuracy at matched compute
mAP is shown in percent. Each RF-DETR variant is paired with the nearest YOLOX variant by parameter count. Across the 4 pairings RF-DETR averages 3.9 mAP points higher and wins 3; YOLOX wins 1. RF-DETR-L tops the range at 58.6 mAP; YOLOX-X, far heavier at 99.07M, reaches 56.3. At the small end RF-DETR-N (30.47M, 51.4 mAP) faces the much lighter YOLOX-Nano (0.91M, 28.8 mAP).
Speed
Averaged across the matched pairs, YOLOX is 32.2% faster than RF-DETR in PyTorch. YOLOX holds high throughput across its whole ladder, while RF-DETR trades frames for accuracy.
- RF-DETR license
- Apache-2.0
- YOLOX license
- Apache-2.0
- Both families
- permissive, cleared for commercial use
Which family to pick
Pick RF-DETR for accuracy per parameter: it wins 3 of the 4 matched pairings and reaches 58.6 mAP. Pick YOLOX for throughput at 32.2% faster on average, and for its smaller variants down to 0.91M parameters. 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.