Skip to main content
Back

Intellindust's DEIMv2 Beats YOLO With 57.8 AP at 50M Parameters

A new detection transformer family plugs DINOv3 features into eight model sizes, with the S variant becoming the first sub-10M model above 50 AP on COCO.

Pro

  • DEIMv2 fuses DINOv3 features with the DEIM detection framework across eight model sizes from Atto to X.
  • DEIMv2-X reaches 57.8 AP on COCO with 50.3M parameters, beating prior 60M+ X-scale detectors.
  • DEIMv2-S is the first sub-10M model above 50 AP on COCO, at 9.7M parameters and 50.9 AP.
  • A new Spatial Tuning Adapter converts DINOv3's single-scale ViT output into the multi-scale features detectors need.
  • Gains come mostly from medium and large objects; small-object AP is roughly unchanged versus DEIM.
  • Checkpoints, ONNX and TensorRT export, and Hugging Face wrappers are shipped in the release.

Object detection has spent the last few years in a two-horse race between YOLO-style single-shot networks and DETR-style transformers. DEIMv2, from Intellindust AI Lab, plants a flag firmly on the transformer side by wiring Meta's DINOv3 vision features into the DEIM training pipeline and shipping the result as an open-source repo with eight model sizes, pretrained checkpoints, and TensorRT export tools.

The paper claims state of the art on COCO across the accuracy-cost curve, and the numbers are striking. DEIMv2-X reaches 57.8 AP with only 50.3 million parameters, surpassing prior X-scale models that require over 60 million parameters for just 56.5 AP. On the small end, DEIMv2-S becomes the first sub-10 million model (9.71 million) to exceed the 50 AP milestone on COCO, landing at 50.9 AP.

DEIMv2 COCO AP versus parameter count compared to DEIM, D-FINE, YOLOv11, YOLOv12

One family, eight budgets

The lineup covers a wider range than most detector releases. DEIMv2 spans eight sizes from X down to Atto, targeting GPU, edge, and mobile deployment, and the backbone choice splits along that axis. The X, L, M, and S variants use DINOv3-pretrained or distilled backbones paired with a Spatial Tuning Adapter (STA), which converts DINOv3's single-scale output into multi-scale features and adds the fine-grained detail that pure ViT features tend to blur. For the ultra-lightweight tier (Nano, Pico, Femto, and Atto), the authors reach for HGNetv2 with depth and width pruning to meet strict resource budgets.

Here is how the official model zoo lays out on COCO val2017:

VariantParamsGFLOPsLatency (ms)