CLIP (Contrastive Language–Image Pretraining)

CLIP, short for Contrastive Language–Image Pretraining, learns to connect visual content with text by comparing many image-caption pairs. During training, matching images and descriptions are pulled closer in a shared representation space while mismatched pairs are pushed apart. This allows the resulting model to classify images from natural-language labels, rank image-text similarity, and support multimodal search or generation systems without task-specific training for every category. CLIP-style models are flexible, but they can inherit cultural bias, weak factual associations, and gaps from web-scale data. Their similarity scores indicate learned correspondence, not guaranteed truth or a complete understanding of the depicted scene.