Prefix conditioning is a novel method for pre-training visual language models using both classification and caption datasets to provide their complementary benefits. Learn how it uses prefix tokens to disentangle dataset biases from visual concepts ↓

→ View original post on X — @googleai