Higher Deviation from the Starting Image

Nobody has claimed this yet.

Thank you for this project!

One of the main limitations I've noticed is that it's difficult to deviate significantly from the starting image — the final frames tend to stay very close to it. I suspect this is why most examples use a static camera. While it's possible to prompt some camera movement, the motion is usually very subtle, and you can't really pan or move the camera to reveal new areas or objects.

It would be great to have an additional parameter (if possible) that allows for greater deviation from the initial image, even if it comes at the cost of consistency.

The issue does not name files, tests, or an entry point. Start by tracing the repository's video-generation configuration and conditioning flow, then determine where a deviation control could be defined and how its effect would be measured; done means users can produce substantially different camera motion or newly revealed content without losing acceptable consistency.

Written by the indexing model from the issue text.