Incidentally, although the image resolution processing is stunning, the motion accommodation 'supposedly' implemented in the 60fps frame-rate conversion is just awful, full of precisely the kind of herky-jerky transitions we decry in older film clips. This is evident as early as the first couple of seconds (with the baggage cart) and is painfully evident in the almost stop-motion progress of the train past the camera.

We had effective motion-vector steering algorithms to remove 'tweening' artifacts in 24-to72fps upconversion (in part developed for 1080P24 'source material' as a broadcast alternative for HD 'back in the day' before the ATSC standard was misadopted for DTV). This kind of effect was very easily removed at 72fps (which was 'flicker-proof' on older short-phosphor CRT displays/monitors). It would be trivial to recode some of the proprietary algorithms even from that era to work with 16fps instead of 24fps (perhaps frame-quadrupled to over 60Hz on modern displays) and to detect and compensate for the variable frame rate introduced by hand-cranking.

Much of the 'focus group' research conducted during the work on TV Anytime clearly established that motion artifacts were of far more concern to HDTV viewers than high resolution, or even particularly accurate 'focus' resolution. I find it a little alarming that this appears to have been so thoroughly forgotten or ignored... especially since the article expresses the idea that the explicit 60fps conversion that has been done is a sufficient technical 'improvement'.

What they don’t tell you is the frame rate of the original film. The slower the original frame rate, the more image interpolation between the original frames you’ll need to get to 60fps. At a slow initial frame rate, you’re asking AI to fill in a pretty big gap between existing frames. And you’re gonna get image artifacts which will be worse with low resolution original images. At some point, you’ll hit the point of negative returns.

That said, if you had a couple wheel barrows full of large denominations, you could probably convert it to full vector CGI at a very high frame rate, but then it would probably look a bit artificial in texture.

Without said wheel barrows, I think they did a nice job.

They don't need to. Everybody knows the Lumiere frame rate is 16fps, and that this is variable in 'reality' due to the issues with hand cranking.

This is little more than a tautology. It is also very well understood by, for example, folks in SMPTE, The point is that computerized 'tweening' has been a remarkably well implemented technology since the early 1990s. If you are not concerned with high resolution precision in the in-betweened frames ... and, if you think about it, when you upscale 'digitally' you're essentially creating a whole lot of utterly synthetic pixel 'precision' ... all you really care about is eliminating the perceived motion artifacts, and there are very well established algorithms for those.

Of course it's computationally intensive to do all the necessary tweening for all the motion pairs between frames at this high nominal resolution. The point to remember here is that even cheap arrays of ARM cores are radically faster than even purpose-built video hardware of previous generations.

The "AI" isn't necessary for most of the motion-vector steering. Recognition of the appropriate regions in the keyframes ... and here, each original frame is a keyframe ... is most of the work involved, and there are tools that easily allow selection and correction of regions in the individual frames if the programs themselves pick wrong.

It may actually not have occurred to you by now that the "low resolution original images" have already been upsampled and smoothed, in the first part of this 'restoration', so the motion-vector steering is working on what are essentially higher-resolution images to start. So this is really a non-issue for the technical motion processing.

Any 'image artifacts' detected in the tweened frames can be quickly identified by nothing more complicated than blink technique ... and hand-edited out as desired.

I think you're not recognizing that typical film 24fps is only one-third greater than what the Lumieres were using, and every bit of the 1080p24 'revolution' is still applicable to film-to-digital transfer. In my humble opinion the 'correct' solution for digital projection is still triple upconversion (to an effective 72fps in the frame buffers) but in part that is because I'm a high myope and even 3:2 pulldown can be violently juddery for me to watch if there is any significant movement in the images between frames.

I repeat that there is no excuse for calling something a 60Hz upconversion and still having jerky motion in the rendered movie. Perhaps he will spend a few extra hours and fix this in the next version.

Interesting. Seriously. As an admitted hard core information junkie, it tickles my synapses since I’ve done image restorations and a little basic video work, both very time consuming. So I jumped down the internet rabbit hole. After looking at the well done examples provided by the DAIN software developers, yeah I agree a little more time could have been spent on smoothing. Maybe the judder was caused by the variable hand cranking rate of the original and the video was done to simply show the potential of the software. Who knows? You’d have to ask the guy that did the conversion why it wasn’t as smooth as the developer’s samples.

But I think a point is being missed: Cost and manhours. The Topaz Labs image software is $99 and seems to be an excellent tool compared to others I do use. I haven’t tried it yet but may as time allows. I would guess that most of the manhours were spent on each individual frame restoration before reassembling and converting the frame rate. The DAIN software is open source. Hardware isn’t that expensive anymore if you need it or you can custom build to suit if that’s your thing. So if you have the time, restoring old films once converted to video really isn’t out of the realm of possibilities for an old nerd like me. The big difference for most of us here is that you’re using a mouse instead of a maul.

Back down the rabbit hole....

Someone has colourised that original French film clip as well now....
https://www.youtube.com/watch?v=EqbOhqXHL7E

The comments section yielded a link to a more current picture of a model of the locomotive. No kidding, it's green... LOL