Lumière’s Train in 4K
thehistoryblog.com
thehistoryblog.com
But (and the 'but' is important here) according to this article, this is not a 4k upscale of a high-quality restoration of the original film, it's based off the second Youtube video on the linked page, which is a terrible low-quality scan of a bad print, haphazardly 'restored' to remove most of the scratches and holes, then uploaded to youtube and compressed to the point that almost no detail remains.
Given the incredibly poor quality of the source material, this is actually pretty impressive.
To me, it creates this eerie kind of feeling where everything looks real-ish, but still has that deceptive feeling of actually being fake.
Maybe I'm just overly exposed to Hollywood CGI and have a good filter for these things?
The original looks low-quality, but real, and so it is clearly better.
I think that's key when it comes to AI for these use-cases.
It's not restoring by filtering out noise and interpolating. It's really creating something entirely new.
Even if they perfect the technology so that the result looks perfectly real, it will still be completely fake.
Maybe even the "weaving" motion is from a slightly swaying cloth projection screen?
I'd be interested to see what the Lumiere train movie would look like if properly laser scanned in 4k with a wetgate.
And it's most probable that the previous kinds of processing were with bad tools or with some other deficiencies.
The proper path would indeed be somebody hi-res scanning frame by frame, all grain and scratches kept, each frame more than once with different lighting etc, and then starting from there and comparing the "restoration" results, always with the original(s).
Just as an example, a lot of techniques for removing grain remove the existing information. The different lighting scans allow a lot of advanced algorithms to restore what our eyes could discern from the real film but most of the processes remove etc.
For me, YMMV, the digitally-restored version still triggers that classification in my head. However, the upscaled one does not. It looks like normal people boarding a normal train, except in black and white. It's really the frame rate that does it for me. This allows a more visceral connection to a time period, and allows us to replace our mental image of "goofy people who are probably just about to drive the train off a cliff or something" with real human beings from a long time ago, doing something mundane and normal.
It isn't there yet, but it's a promising start. Add a decent, realistic colorization and we'll really be cooking with fire.
YMMV, because you may not have this subconscious associations with oldtimey footage or you may be triggering on cues still present in the new footage and still classifying it mentally as "old timey", so you don't see any change. For instance, there's still a very large cue coming out of "black and white video" that may result in your subconscious still classifying it as "old timey" for you. But for those who do have our subconscious classify it differently, as I do, it's an impressive effect.
None of this is meant to be critical of you personally, merely an explanation of why it may be that some people are having a different reaction than you. People are different. I will also criticize myself and say that consciously, I'm well aware that my mental model of "old timey" is completely broken and they were real people back then, just like today, but being consciously aware doesn't affect my feelings anywhere near as much as seeing the upscaled footage does. Humans gonna human.
The tech will only get better.
https://www.amazon.fr/LUMIERE-%C3%89dition-Prestige-blu-ray/...
This is a fantastic blog, and I'd be happy to sponsor/fix its hosting (I run a web-hosting company) - if the writers behind this site read this, feel free to reach out!
I can only imagine the amount of input data and processing power required to do that, though. Probably a decade or two away.
These exist in crude forms because they're considerably more challenging to program than temporal noise reduction algorithms.
Problem #1 is automatic image registration/alignment which is still largely an open problem of great interest to astronomers and those in medical/surgical fields.
It's challenging because we're taking 2D images of objects moving in a 3D space.
Problem #2 is classification. It's very hard at the moment to suss out which high frequency details are important vs. which are noise.