Is there any pretrained dinov2 model with patch size of 16?

[nagitam]

[qasfb]

Not at the moment; does it work for you if you resize your image before the patch embedding block, by a ratio of 14/16 ? I expect it should give the result you want

[nagitam]

Thanks. That was what I was thinking. The use-case is for semantic segmentation and I would have preferred to avoid rescaling.

[ZichengDuan]

Got the same issue. Since as described in the paper, when linear probing the network for semantic segmentation, the patch size is 16 and the image size is 512. Even though I obtained similar results using image size 518 and patch size 14, using the code I implemented from draft.