DINOv2 vs DINOv3: vision features and Gram anchoring

This Korean-language DMQA seminar compares DINOv2 and DINOv3 as self-supervised vision foundation models. Recorded on October 31, 2025 and uploaded on November 23, 2025, it discusses their research methods and reported results rather than installation or hardware requirements.

The speaker explains how an image encoder learns representations from unlabeled images, then supports downstream tasks through separately trained heads. The DINOv2 section covers data curation, DINO loss for consistency between image crops, and iBOT loss for learning masked patch representations. Teacher updates use an exponential moving average of the student. Feature visualizations illustrate similarities between parts of different objects.

The DINOv3 discussion focuses on scaling and Gram anchoring. The speaker describes a 7-billion-parameter vision transformer and explains a limitation of extended training: classification performance can hold up while dense features lose detail. Gram anchoring uses a fixed teacher from an earlier checkpoint and encourages the student to preserve similarities between patch embeddings.

The seminar presents segmentation, depth estimation, object detection and video tracking results. The speaker reports stronger DINOv3 results in the comparisons shown, including detection with a frozen image encoder and a trained detection head. These are the seminar's reported experimental findings, not guarantees for every dataset or deployment.