Visualise Embeddings

3-D PCA projection of DINOv3 image embeddings

Ready

How to read this PCA map

Non-finetuned

PCA projects high-dimensional embeddings into 3 axes (PC1, PC2, PC3). Nearby points share similar visual semantics. Finetuned and non-finetuned variants perform similarly here — on Stanford Online Products the non-finetuned layout is often more separated.

Non-finetuned

General-purpose pretraining features. Performs on par with finetuned and often better-grouped on Stanford.

Finetuned

Dataset-adapted features. Gains are small; class structure is usually similar to the non-finetuned projection.

Base · ViT-B

86M params · 768d · patch 16 · 12 heads · 4 register tokens · RoPE.

Small · ViT-S

21M params · 384d · patch 16 · 6 heads · 4 register tokens · RoPE.

No projection data yet.

Embed your own image