DINOv2 Shows Weaker kNN Performance

psy_com · reddit · 2026-07-08

Testing a frozen encoder with weighted k-NN on a fine-grained car model classification task yielded the following accuracies: SigLIP2 SO400M at roughly 92%, CLIP ViT-L at 59%, and DINOv2 Giant at 41%. The author suggests this gap might stem from different training objectives and asks whether DINOv2 is better suited for linear probes or if switching layers or pooling methods is necessary.

Original post →

More from Models

Models channel →