Scene 3 — the street, and the layout map trick

Scene 3 is the hard one. It's got the most dancing in the whole ad, and the most that has to stay locked frame to frame — the street corner, the backpack, the hero's walking direction, and a giant inflatable skydancer that has no business holding still.

The starting prompt breaks three ways

The first pass exposes exactly what a plain prompt can't hold onto. The light drifts toward sunset when the whole ad is one continuous midday. The hero walks a different direction every generation. And the skydancer is the worst offender — huge in one take, tiny in the next, jumping to a different spot on the sidewalk each time.

Words can't pin down a location — a map can

Text can't fix a spatial problem, so instead of writing more description, a schematic gets built in GPT Image 2.0: mark the fire hydrant, lock the skydancer to its right at exactly twice a person's height, on the same line. That map goes straight to Claude to rewrite the street prompt around it — bright midday, overhead sun, short hard shadows, hero moving left to right with the road on his left and the building on his right.

It's the single biggest trick in this video. Brute-forcing twenty generations to fish out a few usable seconds burns credits and still leaves the geometry loose. A map skips the lottery — showing the model exactly where things sit and how big they are gets Seedance right take after take.

Anchoring the hero, then choreographing every cut

The schematic fixes the skydancer, but the hero still drifts around the frame between takes. The fix is a physical anchor: pin him under a tree on the left, then describe the walkthrough cut by cut with Claude, right down to locking the backpack on both shoulders in every single cut — a detail that had been slipping in and out between takes.

Even with the geography locked, the dancing itself comes out generic and mushy — the same lesson as the kitchen exit: "he dances" isn't a direction. So each cut gets choreographed by name — a slow head-bob, sharp footwork, a locking pose, then a full hip-hop break once he spots the skydancer. Dropping the actual music track into the Seedance prompt as an input locks every move to the beat, and the final cut drops the camera low so the skydancer towers over his shoulder as he answers the call.

Three passes on one prompt — not three separate prompts — and the skydancer holds its size and spot in every single cut.