Option to delay video for more accurate background removal

[pd431]

Issue I face
Using both MediaPipe and RVM, the background removal is fine when I am static, but feels like it is calculated a few frame too late when moving.
This means if I'm moving my head to the right, the rightmost portion of my head is cut out, and the background left of my head is visible.
This happens on CPU or GPU accelerated removal. The effect can be lessened by setting Temporal smoothing to 0, but that means the contour may flicker a lot more than if set to 0.1

Proposed solution

As an alternative to fiddling with temporal smoothing, I would prefer giving a few frames of delay to the engine to allow more accurate matting to be applied. Given the models do a really decent job statically, I wonder whether giving it 2 extra frames to work with might be a possible solution

Possible alternatives

Switching temporal smoothing to zero makes this a lot better (although doesn't fix the issue entirely). Getting a higher quality camera where temporal smoothing set to zero is not quite as flickery might also help.

Additional context
Running Win10, Inference running on Intel Arc 770LE GPU