When views are compressed into latent space, the target embedding naturally removes all noise and contains only compressed abstraction and semantics, allowing the predictor to model only stable structure between views without noise interference.