MiniGPT-4

MiniGPT-4 aligns a frozen BLIP-2 visual encoder with a frozen large language model (Vicuna) through a single projection layer, unlocking GPT-4-like visual conversation from a lightweight training recipe. Presented at ICLR 2024, it became the 25.7K★ ancestor of the entire Mohamed Elhoseiny / Vision-CAIR line — MiniGPT4-video / Goldfish, MiniGPT-Med, and Tempo.

Because it is built on Vicuna plus BLIP-2 rather than pretrained from scratch, its scale is inherited from those bases; the contribution is the alignment recipe and demonstration that a single trainable projection can bridge vision and language for open-ended dialogue.