The clip above starts with a handful of people in frame. In the result, the crowd fills more of the shot, as if the scene always had more people in it. Nothing in the original footage moves; what changed is how much of the scene exists around it.
That's the family of jobs people mean by "extend video with AI": widen a frame for a different aspect ratio, or expand what's happening in a shot. Both ask a model to invent new area beyond what the camera actually recorded.
Why this is different from cropping
Cropping throws footage away. Extending adds it. A horizontal interview cropped to vertical loses whoever is sitting at the edges of frame. Extended to vertical, the top and bottom gain new area instead, so the original subject stays whole in the middle.
The catch is that "new area" is a guess. It has to:
- Match the lighting and film look of the real footage next to it.
- Hold together across every frame, not just invent a static border.
- Stay still when the camera is still, and move when it moves.
Get any of those wrong and the seam between real and invented footage is obvious, usually as a soft, flat patch that doesn't quite match the grain or light of the rest of the shot.
A workflow that holds up
1. Decide what you're actually extending
Converting aspect ratio and expanding content inside the existing frame are different briefs. Widening for 9:16 only needs new area above and below; expanding a crowd or a scene needs new area that still relates to what's already there. Be specific about which one you're asking for.
2. Keep the real footage untouched
The original frame should stay exactly as filmed. The extension sits around it, not over it. If a pass changes anything inside the original frame, that's a sign the request was too broad.
3. Check the seam at full size
Play the result at 100% along the edge where real footage meets the extension. This is where mismatched grain, a wrong shadow direction, or a hard line between old and new pixels shows up first. It won't be visible at a thumbnail size.
4. Keep extensions modest near busy edges
A plain sky or a quiet background extends cleanly. A shot with people, signage or motion right at the edge of frame is harder, because the model has to guess what those things do as they continue past where the camera could see. Expect to review those more carefully, or trim the shot so the busy part isn't right at the edge.
Asking an agent to do it
Extend shot 2 upward and downward for a 9:16 crop. Keep everything in the original frame exactly as it is, and match the lighting of the new area to the rest of the shot. Show me the result before replacing the original.
With the run mode set to review, the agent generates the extension and waits for approval before anything replaces the original clip. You still make the call on whether the seam is clean enough to ship.
Where this still needs a human
- Judging the seam. Whether an extended edge reads as part of the shot or as a visible patch is a eye-test call, not something to automate away.
- Busy edges. The more is happening right at the frame line, the more the extension is guessing. Review those shots frame by frame rather than trusting a quick preview.
- Repeated extensions. Extending the same shot more than once compounds the guesswork. Start from the original footage each time rather than extending an already-extended result.
Download the free Shoty editor to try it on your own footage. Related: how to remove an object from video with AI.