Use Inpainting for Consistency on Single Prompt Animations

[mikayahlevi]

I believe that Deforum videos of a single prompt would be more consistent if there was a mode that instead of generating a new frame, inpainted the missing details. For example for a 2D camera that is zooming in what it would do is keep all of the image the exact same (maybe with noise scheduling settings) except for the center which it would inpaint at a strength proportional to the zooming speed, so that the zoomed area would be in proper resolution. For a 3D rotating camera as another example, it would find the areas that MiDaS shows weren't in view last frame.

[kabachuha]

Looks similar to my pull-request of #74 (there're issues with noise and we together with @Funofabot are thinking on a more elegant solution)

[Funofabot]

We need to revisit this. Over the weekend hopefully.

[mikayahlevi]

This is actually almost exactly what I was imagining:
https://ai.googleblog.com/2022/11/infinite-nature-generating-3d.html
it could be a separate mode or maybe a schedule.
Also, when parts get too zoomed in and blurry maybe it could be upscaled?