The original image diffusion models used a U-Net to predict noise. Replacing it with a transformer changed the scaling behaviour: performance improved predictably with size and data in a way the convolutional version did not, which is the reason the current generation of video models exists at all.
For a producer, the practical consequence is that quality now tracks compute more reliably than it tracks clever prompting. The gains of the last two years came from bigger models trained longer, not from a trick anyone can apply at the keyboard.
The second consequence is cost structure. Transformer attention scales badly with sequence length, and a video is a long sequence, which is why clip length is priced the way it is.
Why did video generation improve so quickly?
Largely because the denoising network became a transformer, which scales predictably with compute and data. Most of the visible progress is scale rather than a technique that transfers to prompting.
Does this matter when writing a prompt?
Not directly. It matters when reading a pricing page, because attention cost grows with sequence length and that is why longer clips cost disproportionately more.