There is a quiet argument happening inside every text-to-video demo. One side says: give the model the most freedom you can, and it will surprise you. The other side says: freedom is where the artifacts live.

The scream-and-swap dance trend — two photos in, one 21-second vertical clip out — has already picked a side, and I think it picked correctly. It gives the model almost nothing to improvise. That is not a shortcut. It is the entire reason the output works.

Free-form generation is a reliability tax

When you ask a video diffusion model for "two people dancing together," you are implicitly asking it to solve several hard problems at once: consistent identity for two distinct subjects, plausible motion that respects a beat, spatial layout that keeps two bodies from merging, and temporal coherence across every frame. Each of those problems compounds the others.…