inference4j/dfine-nano-coco

D-FINE Nano (COCO) — ONNX

ONNX export of ustc-community/dfine-nano-coco, the nano variant of D-FINE (Redefine Bounding Box Regression for DETR-based Object Detection) trained on COCO (80 classes) with static shapes for full ONNX Runtime and CoreML compatibility. Licensed under Apache-2.0, providing a commercially permissive, non-copyleft alternative to AGPL-licensed YOLOv8 and YOLO26 models.

Mirrored for use with inference4j, an inference-only AI library for Java.

Original Source

• Repository: USTC / onnx-community (D-FINE)
• License: apache-2.0

Usage with inference4j

try (DFine model = DFine.fromPretrained("inference4j/dfine-nano-coco")) {
    List<Detection> detections = model.detect(Path.of("street.jpg"));
    detections.forEach(d -> System.out.printf("%s (%.0f%%) at [%.0f, %.0f, %.0f, %.0f]%n",
        d.label(), d.confidence() * 100,
        d.box().x1(), d.box().y1(), d.box().x2(), d.box().y2()));
}

Model Details

Property	Value
Architecture	D-FINE Nano (transformer object detector)
Task	Object detection (COCO 80 classes)
License note	Apache-2.0 (permissive non-AGPL alternative to YOLOv8/YOLO26)
Input	pixel_values: [1, 3, 640, 640] — RGB, normalized 0-1 (bilinear resize, no letterbox, static shapes for CoreML)
Output	logits: [1, 300, 80] (raw logits, apply sigmoid) and pred_boxes: [1, 300, 4] (normalized [cx, cy, w, h])
Post-processing	Convert [cx, cy, w, h] to [x1, y1, x2, y2] scaled to image dimensions; filter by score threshold (> 0.5). No NMS required.
Hardware support	Verified on CPU and Apple Silicon CoreML (NeuralNetwork & MLProgram formats).
Original framework	PyTorch (USTC D-FINE)

License

This model is licensed under the Apache License 2.0. Original model by USTC / Peng et al., hosted by ustc-community and converted by onnx-community.