IC-646DINOv2, DeiT-III, and OpenCLIP repurpose approximately 2% of patch tokens in low-informative background areas as internal registers, discarding local patch information while aggregating global image information; DINO does not exhibit this behaviour

Timothée Darcet, Maxime Oquab, Julien Mairal, Piotr Bojanowski

SourceVision Transformers Need Registers

The paper identifies high-norm outlier tokens (norm > 150) in the output feature maps of DINOv2, DeiT-III, and OpenCLIP. These tokens correspond to patches that are highly similar to their four neighbours (cosine similarity 22.8 vs 41.7 for normal patches), indicating redundant background content. Linear probing shows these outlier tokens retain far less local information: position prediction accuracy drops from 41.7% to 22.8% and pixel reconstruction L2 error rises from 18.38 to 25.23. Conversely, using a single outlier token as the image representation for classification yields dramatically higher accuracy than a normal patch (e.g. 79.1% vs 17.1% on Aircrafts), indicating they store global image information. The artifacts emerge around layer 15 of a 40-layer ViT, appear only after roughly one-third of training, and are present exclusively in models of ViT-Large size or larger. DINO ViT-B/16 shows no such outliers, and MAE ViT-Large also lacks them.

Evidence
correlational
Key metric
2.37% of DINOv2 ViT-G/14 tokens have norm > 150; position prediction top-1 acc: normal 41.7% vs outlier 22.8%; reconstruction L2 error: normal 18.38 vs outlier 25.23; cosine similarity to 4 neighbours: normal 41.7 vs outlier 22.8; Aircrafts linear probing top-1: normal 17.1% vs outlier 79.1% vs [cls] 87.3%; artifacts appear around layer 15 of 40; only in ViT-Large, ViT-Huge, ViT-Giant
Caveat
The norm cutoff of 150 is hand-picked and can vary across models. The authors note they have not fully determined which aspects of training (pretraining paradigm, model size, training length) cause the artifacts to appear.
Model
DINOv2, DINO, DeiT-III, OpenCLIP, MAE
Concepts
Register tokens, Depth-dependent structure, Scale-dependent behaviour
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], ImageNet-21k / ImageNet-22k [eval]
Methods
Linear Probing / Ridge regression linear probing / Linear probe / Linear probe fine-tuning / Linear regression probing / Linear ridge regression probes / Supervised probing / ERM linear probe [primary], Cosine similarity / Cosine similarity analysis / Cosine semantic similarity / cosine similarity of hidden states / Sample-wise cosine similarity / Cosine similarity of attention maps / Cosine similarity perturbation analysis / Cosine similarity template matching / Cosine similarity to neighbours / Semantic consistency (cosine similarity) [primary]
Related work
Memory Transformer [context]
Related findings
IC-647
Extraction
automatic-extraction