tepples wrote:
The question then becomes one of whether subjective visual improvement due to right guesses outweighs artifacts due to wrong guesses.
No, that's not the question. There is no "right guess"; there is no data there to be guessed at all.
It a question of whether the fictional additions to the image are acceptable, or even necessary at all.
Even if you were trying to tune your algorithm by comparing a downsampled version against a multisample source, the closest match to the original is not really the right metric to use. There's a lot of tradeoffs involved when thinking also about temporal coherence, etc. but with upscalers like this the goal is
not trying to recover the original picture. It's about trying to come up with a
plausible lie.
This is not a signal recovery algorithm. That's the most irksome part of Marseille's marketing to me, where they start talking about presenting it "as the director originally intended", "restoring depth of field", "preserved textures", etc.