Hacker News Reader

Comment by imiric | Hacker News Reader

I wonder what kind of results modern upscaling software could achieve with that less detailed scan. I've seen much more impressive improvements than what it would take in this case, and I reckon the end result could look even better than the Hasselblad scan.

I'm sure film purists would scoff at this, but there's a large market that would prefer cheaper solutions that do this in software. We're already used to these enhancements in modern cameras, so why not scanners?

View on HN

That's just making up stuff, though. I don't think only "film purists" scoff at this. Anything that even remotely has the chance to be used for archival purposes must not be upscaled that way. Not only do you effectively fill in details that simply aren't there, you also don't even have any foolproof way to know which details are made up by upscaling. Worse, some super fancy upscaling might falsify parts of the image that are still above the threshold of where you would need upscaling.

Imagine a historian going through old photos (potentially ones that were previously deemed uninteresting, even) and is led astray by made up details. Imagine a criminal investigation doing the same.

In short, unless you have very specific purposes in mind, and always be careful to clearly mark upscaled images as such and to keep the originals, upscaling film photos that way is a terrible idea.

Before anyone asks, I do think there is some value to upscaling of old photos and videos, like the popular HD/4k 60fps upscaling of videos from the 1800s: It gives even casual viewers a better sense of what the actual scene, and the life depicted within, might have been at that period. It's easier to accept this as "more real" (even though it's ironically less real).

But I really wish any such upscaling was very clearly marked as such, and pointed out to even casual viewers with disclaimers: Unlike the old choppy, blurry, low-res video, this is not a sample of reality back then anymore, but a speculative reconstruction.

To be fair upscaling with videos isn't necessarily making information up, but depending on the method used may use information from other frames, wich I think isn't the same as speculative reconstruction for photos. For increasing the frame rate for videos this depends on the model used, but in general most of them don't add any new information, but instead calculate how something looked in-between two frames. https://medium.com/axinc-ai/flavr-a-machine-learning-model-t...

Definitely! I was explicitly talking about things like "AI upscaling". Not simple interpolation (which keeps the image spectrum the same and is effectively just anti-aliasing), for example.

To be honest though, the method you linked does seem to fall in the bucket I am warning against: It does make information up, it does add that made up information. Those frames in between are not simply interpolated, or you would not need an ML model.

Case in point, individual droplets in that example animation may not have existed in reality, or moved the way they did here. But hopefully that's fixed by just dropping the interpolated frames, at least.

> That's just making up stuff, though.

It's taking the next step in line with what all digital cameras already do. They all do post-processing and make approximations to what the eye actually sees. Just because upscaling technology is still in its infancy and it's not entirely foolproof yet, is not an indication that it won't become seamless and undetectable in the future.

I agree that such features should be optional and all processing should be clearly marked, but to dismiss the technology altogether is a mistake.

Those upscalers never looks remotely correct to the trained eye beyond minor scaling increases, even with state of the art (I work in post production).

Reply on news.ycombinator.com