From CLIP to BLIP-2 — Q-Former as Modality Bridge

CLIP aligns image and text but cannot generate captions, answer questions, or hold a conversation. BLIP-2 (Salesforce, 2023) solved that with a small trainable bridge: 32 learnable query vectors attend over a frozen ViT's features via cross-attention, then slot directly into a frozen LLM's input stream. 188M parameters of bridge connected an 11B LLM to a ViT-g/14. Every adapter-based VLM through 2026 — MiniGPT-4, InstructBLIP, LLaVA's cousins — is a descendant. This lesson reads the Q-Former's architecture, explains its two-stage training, and builds a toy version that feeds visual tokens into a frozen text decoder.