Four ways of feeding the timestep and the text prompt into a diffusion transformer block.

The block has no slot for “what to draw”

A transformer block takes a sequence and returns a sequence. But this one also needs to know how noisy its input is, and what the picture is meant to be of — and neither of those is a visual token. So where do they go in?

A label is one thing; a sentence is a sequence. Three of these squeeze the condition into a single vector. The fourth stops trying, and gives the words a tower of their own.