[kasool]

This has long been the aspect I've struggled with, having spent some time implementing various temporal solutions in the past 2 years. We can try and pull out all sorts of classic metrics like SSIM, but the truth is a lot of these effects are really hard to objectively evaluate using some sort of metric. Moreover the same technique can have vastly different outcomes depending on the content, and user perception is subjective as well. Many days were spent thinking I had solved a visual issue, only for another edge case to come up under specific conditions. Many of these techniques are fairly difficult to reason about from first-principles and adding ML into the mix only makes that harder.