Open Nano Banana Pro (gemini-3-pro-image) and ask for a real photograph, not a render:
For the stadium: upper tier behind one goal, night, floodlights, packed stands, 24mm, sensor noise in the dark. Set the aspect on the node (4:3 here, so it fills half of a 9:16 reel).
Real places work: name the actual landmarks. This still is the only location the model is allowed to fly through.
Draw a red line with an arrow over the still: where the camera starts, what it dives over, where it exits. Any editor works, even a finger on a phone.
The moves that read best:
Keep the original still too. You will use both.
In Seedance 2.5, wire BOTH stills into reference_images: the clean still first, the marked still second (the model refuses references together with a start frame, so do not use one). 4:3, 1080p, 10 seconds, audio on.
Write the prompt so it tells the model the first reference is the opening view, the second carries the path, and the red line must never appear. Audio on gives you crowd, wind and floodlight hum for free.
Keep the sentence "fly only through the exact place shown in the references, do not invent landmarks". Without it the model drifts to a more famous version of your location.
It rendered Budapest instead of Bratislava on my first city flight. If it still drifts, pass the clean still twice and describe the real landmarks by name.
Both models ran inside Claude Code through the ElevenLabs connector (add it with claude mcp add --transport http elevenlabs https://api.elevenlabs.io/v1/mcp), one login for the still and the flight, and you can ask for the credit estimate before the run.
Then post the comparison, not just the flight: the marked still on the bottom, the video on top, each with its model name. The drawing is what makes people stop, and for a client the red line is the storyboard.
This guide is the full process. The ready-made version lives inside AI Creative OS on Skool, next to every other prompt pack I post.
