flamingo

Definition

A vision-language model from DeepMind that accepts interleaved sequences of images and text and can be prompted few-shot, the way a language model is, on tasks it was not trained for. Its architecture keeps a pretrained vision encoder and a pretrained language model frozen and trains only the cross-attention layers that bridge them. That bolt-the-eyes-onto-an-LLM pattern is the direct ancestor of most later multimodal systems.