Facebook DETR Alternative in 2026: Run It with LibreYOLO
Xuban
The original Facebook Research DETR repository was archived on March 12, 2024. LibreYOLO keeps the original detector usable through the same API as its other model families.
from libreyolo import LibreYOLO
model = LibreYOLO("LibreDETRr50.pt")
result = model("image.jpg", save=True)
print(result.boxes.xyxy)
Four ResNet-50 and ResNet-101 variants are available for prediction, validation, ONNX and TorchScript export. Training is not implemented. LibreYOLO also uses a fixed 800 by 800 input, rather than reproducing the original aspect-preserving evaluation pipeline.
Try it after pip install libreyolo. The DETR docs list all four checkpoints. For a CNN comparison, see the EfficientDet alternative.