Alongside the per-token sequence, many systems also produce a pooled vector — a single summary of the whole prompt, often taken from a special start-of-sequence token or by averaging. The pooled vector is cheap and captures overall gist, which is useful for global control. But it cannot tell the generator which region of the frame should contain the boat and which should contain the rain. The per-token sequence preserves that binding; the pooled vector throws it away.
How Text-to-Video AI Works Under the Hood
From Prompt to Conditioning Signal
Sequence Versus Summary
3 / 4
Here is a fork in the road. Some systems keep that whole sequence. Others also collapse it into one pooled vector, a single summary of the entire prompt.
0:00 / 0:00