Apple proposes STARFlow2, a unified multimodal model architecture combining normalizing flows with frozen pretrained vision-language models