Not necessarily true. It is QUITE old hat to be able to look at 2+ frames before and after to guess what a frame in the middle should be, this has been around for awhile used to analyze frames to better help noise/temporal noise reduction, I used to use that in megui. There is no way someone can tell me that AI, with how powerful it is now, cannot take a few frames before and after and THEN use AI matching patterns, and if none found just best guess placement between frames. It should easily be able to see there are no squares in that.
Now… that said this is WHY I provided this, because you do not optimize and modify for the BEST case scenarios, you do it for the worst ones because you know if you get the worst down, the best are easy. This is, of course, sans other models which AI is learning, and it very much should have a path of easy reproduction if their standard models do not apply by analyzing multiple frames. I could be wrong 100% on how this works, but I already do this type of thing with our AI bots we have deployed at work, if our trained responses and models do not suffice, they take the context given and try to extrapolate from that.
edit Additionally, I would think it should not be a monumental task to include an artifact hunter addition that essentially views each frames ±5 from the frame it is on to see if that frame contains artifacts, either self generated or from compression, and rebuild as necessary. I say ±5 because a scene change in the middle could throw that off, and I cannot imagine anything that has 2 scene changes in 5 frames. look at 3 frames before look at 3 frames after notice single frame artifacts not in the other original frames rebuild those frames using a frame-to-frame model and not a learned model something like that would make sense to me. Obviously standard is training models, and the more training the better it can “guess” the data, but if you have lines pre and post that frame that move a tad, it should be easy to replicate the middle frame of that. My case IS one of the worst case, so if they get it to work on that, it should be fine for all other light flashes, explosions, electricity jumping scenes, etc. Make it an option for this, make sure all knows this will be for high action/movement movies or scenes and that it will dramatically slow down processing due to the frame compares, but for high speed movies or scenes this could be invaluable…
