Using MiDaS (Depth Estimation)

[Selur]

From time to time, I wonder whether blurring the background slightly for compression gain would be worth it. (Idea would be to use a depth map or an inverted depth map for masking.)
There is MiDaS (https://github.com/HolyWu/vs-midas) which does seem to work in general, but it has a lot of flickering, and I was wondering whether someone here already did play around with it and has some suggestions?

Cu Selur

[Zarxrax]

A model like Video Depth Anything (https://github.com/DepthAnything/Video-Depth-Anything) would probably be better because its specifically designed for video, so shouldn't have the flickering problems. Maybe someone could make a vapoursynth wrapper for that one.

I have tried mitigating the flickering issues before, but aside from something like a temporalsoften, couldn't find anything to really help much.