Feature Atlas

All 12,288 features from the sparse autoencoder trained on DINOv3 ViT-S. Click any tile to see its nearest neighbors.

Companion to Playing with Vision Embeddings.