Pang

Vision encoders share common geometric structure

Researchers have identified a consistent geometric structure, termed the "cross-architecture substrate," within modern vision encoders, regardless of their specific training objective or domain. This substrate, a 16-dim…