Please note: The reported FPS numbers for the DINOv3 models are not directly comparable to those of the DINOv2 models, due to differences in drivers, PyTorch version, and other software updates that were done in the meantime.
FPS measured on NVIDIA H100 (torch.compile default / torch.compile with max-autotune).
| Config | Input size | FPS | PQ | Download |
|---|---|---|---|---|
| EoMT-S2x | 640×640 | 246 / 570 | 47.2 | Model Weights |
| EoMT-B2x | 640×640 | 240 / 347 | 53.1 | Model Weights |
| EoMT-L | 640×640 | 137 / 157 | 56.8 | Model Weights |
| EoMT-L | 1280×1280 | 33 / 34 | 58.9 | Model Weights |
2x Longer training schedule.
| Config | Input size | FPS | mIoU | Download |
|---|---|---|---|---|
| EoMT-L | 512×512 | 137 / 200 | 59.5 | Model Weights |
FPS computed on random 512x512 inputs.
| Config | Input size | FPS | mAP | Download |
|---|---|---|---|---|
| EoMT-L | 640×640 | 133 / 157 | 45.9 | Model Weights |
| EoMT-L | 1280×1280 | 33 / 34 | 49.9 | Model Weights |
Important: The provided model weights are deltas with respect to DINOv3 weights. Users need to obtain access to the original DINOv3 weights first before using these models.