Skip to content

D-FINE

Real-time object detection with refined distributions

Visit official site

What D-FINE is

The model is described as a real-time object detector that redefines DETR bounding box regression and adds self-distillation without extra cost.

What you can do with it

  • Detect objects in real time with high accuracy
  • Pick a model size for the compute budget
  • Use COCO and Objects365 pretrained weights

Application scenarios

  • Track several targets in surveillance video
  • Locate pedestrians and vehicles on a road
  • Help a robot recognise and grasp objects

Pros & cons

✓ What we like

  • Model sizes fit different compute and latency needs
  • Code and pretrained weights are open source

! What to watch out for

  • Detection builds on the DETR regression redesign
  • Pretrained models come from COCO and Objects365

FAQ

What is FDR?

Fine-grained distribution refinement, which iterates probability distributions for precise localization.

Are weights available?

The material states code and pretrained weights are open source.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review