Autoregressive
One token per pass, left to right.
Pass 0 · Settled 0/16Done
Masked diffusion
Every pass guesses all the blanks and keeps the surest ones.
Pass 0 · Settled 0/16Done
Simplified. Both rows advance one forward pass per tick. A diffusion pass reads the whole sequence, so it costs more than one autoregressive step, and real models use more rounds than four.