blip-

Definition

A family of vision-language pretraining models, usually cited with a version number after the hyphen. BLIP-2 is the best known: it keeps both the image encoder and the large language model frozen and trains only a small bridging module, the Q-Former, which extracts a fixed set of query tokens from image features for the LLM to consume. Freezing both ends and learning only the connector cut training cost sharply, and the pattern carried over into much later multimodal work.