Gemini Omni tests sketched camera paths from map screenshots

• bilawalsidhu's original demo showed Gemini Omni turning a hand-drawn path into drone-style POV footage, and bilawalsidhu's follow-up said the first reference was a real 3D scan before he simplified it to a scribble on an Austin map screenshot.
• techhalla's Middle-earth test pushed the same trick beyond real places: one map image, one arrow, and a prompt for a horse-riding tour through fictional terrain.
• The workflow is already mutating into a format, with bilawalsidhu's trend post pointing to a Seedance example and the KEETY repost amplifying another fast follow.
• The current outputs still leak their scaffolding, because MayorKingAI's reply called out visible guide lines and the drone itself in the rendered video.

bilawalsidhu's original demo started with a rough flight line over a map, then bilawalsidhu's follow-up revealed the more interesting detail: the method held up after he swapped a real 3D scan for a quick scribble on a screenshot. techhalla's fantasy-map version suggests the model is reading directional intent, not just matching known geography, and bilawalsidhu's later post shows the idea already hopping to Seedance-style experiments.

The core reveal is simple: a rough line on top of a reference image can act like camera blocking.

According to bilawalsidhu's follow-up, the first pass used camera poses from an actual 3D scan. Once that worked, he reduced the input to a scribbled path on a 3D map screenshot of Austin.

That progression matters more than the demo clip. It turns planning art into usable control, from structured pose data down to a loose sketch.

techhalla's prompt write-up gives the clearest recipe in the evidence set. The setup was a map image plus instructions to read the arrow, remove it, and generate a tour-guide video.

The prompt structure breaks into four parts:

1. Analyze the location and direction of the arrow.
2. Remove the red arrow from the final scene.
3. Visit the two most important locations in that place.
4. Add a POV style, in this case a horse ride with gritty found-footage texture.

Because the example used Middle-earth rather than a real city, the result looks less like route reconstruction and more like semantic camera planning from a marked image.

bilawalsidhu called camera-path scribbles an AI video trend within a day, and he pointed to a Seedance example by Keety rather than treating Gemini Omni as a one-off.

the KEETY repost is thin on explanation but useful as evidence of how fast the format jumped. People are not waiting for a polished product surface here, they are copying the input pattern.

That makes this look closer to a new control primitive than a single viral prompt.

The sharpest critique in the thread was practical. MayorKingAI's reply said the idea worked, but the line and the drone should disappear from the finished shot.

That complaint identifies the current ceiling: the model can follow rough motion cues, but it does not always hide the scaffolding cleanly. For creators, that leaves a distinctive in-between aesthetic, half previs, half finished footage.