I've - just given up on re-installing SVP. Just doesn't work on my machine. It worked last time though, and I think there might be something fishy going on in the code regarding VM-Detection. Or it's just the lack of a physical GPU that's bothering the SVP Manager at startup. Either way it would require me to take a way more complicated approach than I would get with FlowFrames at the moment. Still, it isn't ideal, updating as a non-paying user also takes about half an hour which isn't great either. So I'll leave it be for now.
As for 4.6, I've just tested it with some black and white scenes and though I didn't compare to 4.4, it looks a lot better than v4. B&w seems to be really difficult for interpolation as it doesn't know where that specific dark spot came from, from above or from the side etc., whereas before it could just tell them apart by color.
Back to the original question: Since I think the same AI fed with the same image sequence will always return the same output, maybe cut out that scene, use a different RIFE version und paste it back in? A little complicated, I know, or you could just concede and stick with v4.4 for the whole movie? To me it doesn't look like a technical issue with neither SVP nor RIFE AI, just that the training data is always a bit different and sometimes just gets worse with an update in some areas. And v4.4 isn't vaaastly inferior to v4.6 imo...
Also, I'm well aware this is not a FlowFrames troubleshooting forum, but I'd be happy to help there if needed :)