What is contrastive vision-language pretraining (CLIP-style models)?

CLIP learns joint embeddings for images and text using contrastive learning.

CLIP trains two encoders (image and text) to map inputs into a shared embedding space. It uses contrastive loss to maximize similarity of correct image-text pairs while minimizing similarity of incorrect pairs. This enables zero-shot classification, retrieval, and cross-modal reasoning without task-specific training.