The training involved human writers crafting detailed captions, harmonizing subject and context. This change reduced caption 'noise', enhancing training accuracy. But more importantly, they used these human data to train a captioner model to create synthetic data!

→ View original post on X — @whats_ai