D-FINE
PeterandeD-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- 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.
- https://huggingface.co/api/models?author=ustc-community&search=dfine&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
Thirteen ustc-community D-FINE checkpoints, all gated false and tagged apache-2.0.
- https://huggingface.co/ustc-community/dfine-xlarge-coco/raw/main/README.md recorded 2026-11-12
dfine-xlarge-coco card: license apache-2.0; "This is the HF transformers implementation for D-FINE"; _coco trained on COCO, _obj365 on Objects365, _obj2coco on Objects365 then COCO.
- https://raw.githubusercontent.com/Peterande/D-FINE/master/LICENSE recorded 2026-11-12
LICENSE is the Apache License, Version 2.0.
- https://raw.githubusercontent.com/Peterande/D-FINE/master/README.md recorded 2026-11-12
README: COCO and Objects365 model tables with yml configs and checkpoint links; COCO2017 and Objects365 preparation steps; "License note: Checkpoints trained or pretrained on Objects365 (*_obj365, *_obj2coco) may be subject to the Objects365 dataset terms and should not be assumed to be commercially cleared under the D-FINE license."
Adoption
2 high confidenceHugging 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.
- https://huggingface.co/api/models?author=ustc-community&search=dfine&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
dfine-xlarge-obj365 16,734, dfine-xlarge-coco 12,570, dfine-large-coco 11,102 and dfine-xlarge-obj2coco 8,408 downloads in the trailing 30 days (48,814 together); the thirteen checkpoints sum to 73,796.
Capability
3 high confidenceD-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.
- https://raw.githubusercontent.com/Peterande/D-FINE/master/README.md recorded 2026-11-12
README model zoo: D-FINE-N 42.8 COCO AP; D-FINE-S to X 50.7 to 59.3 COCO AP with Objects365 "Objects365+COCO" pretraining; "Add custom dataset finetuning configs".
Verified 2026-11-12