Pinned
For years, text-to-image generation has scaled models, data, and compute. Yet every training image still reaches the model through a caption, which determines what visual information can be associated with text.
Can caption information itself be measured and scaled?
Introducing
GIF
