Hacker News Reader

Comment by jcynix | Hacker News Reader

While I agree with the gist of your comment, I cannot leave this detail uncommented

>Or, put another way: you can undo effects of a lens, or the way photo was developed classically, because each pixel is still correlated with reality,

You cannot undo each and every effect. Polarizing filters (filters, as are lens coatings, are part of a classical lens in my opinion), gradual filters, etc effectively disturb this correlation.

As does classic development, if you work creatively in the lab (as I did as a hobby a long time ago in analog times) where you decide which photographic paper to use, how to dodge or burn, etc.

But yes, I agree that computational photography offers a different kind of reality distortion.

View on HN

> You cannot undo each and every effect.

Fair enough.

> Polarizing filters

Yeah, I see it. This one is as pure signal removal as it comes in analog world. And they can, indeed, drop significant information - not just reflections, but also e.g. by blacking out computer screens - but they don't introduce fake information either, and lost information could in principle be recovered -- because in reality, everything is correlated with everything else.

> But yes, I agree that computational photography offers a different kind of reality distortion.

A polarizing filter or choice of photographic paper won't make e.g. shadows come out the wrong way. Conversely, if you get handed a photo with wrong shadows, you not only can be sure it was 'shopped, but could use those shadows and other details to infer what was removed from the original photo. If you tried the same trick with computational photograph, your math would not converge. The information in the image is no longer self-consistent.

That's as close as I can come up to describing the difference between the two kinds of reality distortion; there's probably some mathematical framework to classify it better.

The most extreme case being Samsung outright painting the moon in when it thinks a piece of picture is semantically likely to be the moon.

Whether the detail data is encoded as PNG or as AI weights is immaterial, it is adding data that is not there, by a long shot.

https://www.theverge.com/2023/3/13/23637401/samsung-fake-moo...

That's weird. Whenever I tried to take a picture of the moon, it would look great in the camera view on the screem, but look terrible once I actually took the picture.

Reply on news.ycombinator.com