CLIP (Contrastive Language-Image Pretraining) trains a vision encoder and a text encoder jointly on large-scale image-caption pairs using a contrastive objective. For each batch, the model computes cosine similarity between every image embedding and every text embedding, then optimizes so that the similarity between correctly paired image-text embeddings is maximized while similarity between all mismatched pairs in the batch is minimized (an InfoNCE-style loss). The result is a shared embedding space where semantically related images and text land close together, regardless of modality.

This is why CLIP generalizes to zero-shot classification: given a new image and a set of candidate text labels (e.g., "a photo of a dog," "a photo of a cat"), the model simply picks the label whose embedding is closest to the image embedding — no task-specific fine-tuning required. This same mechanism underlies CLIP's role as the text-image alignment backbone in generative pipelines like Stable Diffusion's guidance mechanism.

Options A and C describe mechanisms CLIP does not use — there is no frequency-domain transform or image-to-text translation step — and D describes a static lookup system, which would not generalize beyond its predefined database. Contrastive learning's dual-encoder, shared-embedding-space design is the defining architectural feature to remember.