CLIP

leverages a contrastive learning approach, training separate image and text encoders on a massive dataset of 400 million image-text pairs to predict the most relevant captions for images, enabling impressive zero-shot transfer capabilities to various downstream tasks without requiring task-specific training data.