Back to Gap Map Infrastructure / Classic ML & computer vision

D-FINE

Peterande
open source / Overall score: 2.4

D-FINE is a real-time object detector in the DETR family that recasts bounding-box regression as fine-grained refinement of distributions over box edges, and adds a localization self-distillation step that costs nothing extra at inference. It comes in Nano to X sizes, trained on COCO or pretrained on Objects365, with training code and configs for fine-tuning on custom datasets.

The Hugging Face checkpoints are Transformers conversions published under a University of Science and Technology of China account; the repository's own checkpoints download from GitHub release assets.

Openness

5 medium confidence
5.0
weights
open (Nano to X checkpoints from GitHub release assets and Transformers conversions on the Hub, ungated)
data
open (COCO 2017 and Objects365, both public, with preparation steps in the README)
code
open (training code and configs for COCO, Objects365 and custom datasets)
license
Apache-2.0 (code and checkpoints)
objects365-terms
caveat (the README warns that checkpoints trained on Objects365 may be subject to that dataset's terms)

D-FINE's code and checkpoints are Apache-2.0, the training code and configs are public, and the detectors are trained on COCO and Objects365, both public datasets. The README cautions that checkpoints trained on Objects365 may carry that dataset's terms and should not be assumed cleared for commercial use.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing 30 days for the four most-used Transformers checkpoints. Checkpoints fetched from the repository's GitHub release assets are not counted here.

Capability

3 high confidence
3.0

D-FINE and RT-DETR are the same kind of product, real-time transformer detectors trained on fixed benchmark classes, so detecting anything else needs a training run on the user's own labels.

Verified 2026-11-12