So I'm going to chime in more or less to talk about something more tech and methodology related so sorry if it isn't on topic enough. I've done a bunch of paper reading on the subject with both image and video resolution. It seems to me the main issue is still that when we're talking about video upscaling of anything AI, we're still talking fundamentally about approaches that use image upscaling and apply it to all the frames of a video. But that doesn't take into account the extra information a video provides in terms of temporal information or motion which you can get over several frames. I'm talking about an approach that is similar to what Nvidia does with DLSS but with less information than is afforded by a game engine.
If you look into academic literature for state of the art video upscaling, this is exactly the approach one should be taking. However, the main issue is that nothing is really set up to do this type of thing on even a consumer ready scale. Not that there is anything quite wrong with it but that does seem like a major issue that will impede progress on getting better video output on superresolution, regardless of whatever model or tools used.
That being said, now for the controversial bit because of some ludditism here. I think https://openmodeldb.info/models/2x-AnimeSharpV2-RPLKSR-Sharp is really the new king in terms of speed to quality ratio with anime as PLKSR (https://github.com/dslisleedh/PLKSR) has academically been measured as the most efficient model this year. I know it has problems with depth of field but for me does great and has a soft version which outperforms most Real ESGAN models. However, I am split because he did release https://github.com/Kim2091/Kim2091-Models/releases/tag/2x-AnimeSharpV3 which he claims that is better than v2 in all situations and solves the depth of field issue. While I disagree with the first assertion, the 2nd one is valid so there might be a battle brewing on if RPLKSR can replace Real ESRGAN at last.