Hacker News Reader

Comment by afavour | Hacker News Reader

When I stop to think about it I feel like photography these days in such a weird state. We started with analogue photography that literally captured the incoming light, today we’re taking multiple photos and merging the results together… before long we’re going to be at the point where your camera software recognises you’re stood in front of, say, the White House, goes online, grabs stock photography of the White House and merges it into the shot you took.

I don’t know at what point it stops being a photo any more. In many ways I suppose it doesn’t really matter but I wonder if I’ll be looking at my kids graduation photos and seeing a smile an AI has approximated my child probably had in the moment based on previous photos and stock photography of graduations.

View on HN

> I don’t know at what point it stops being a photo any more.

Personally, I think it's at the point that you start using AI/ML for anything other than the parameters involved in capturing the raw image (focus, shutter speed, aperture, etc) and the parameters of RAW processing/the sorts of editing that's allowed in a typical photography competition (so things like exposure, white balance, colour adjustment, etc).

I'll even allow it to change the exposure/balance/adjustment in parts of the picture separately and it's still a photo. Once you cross that Rubicon though I think it's a different thing.

I think the analysis can get a bit more blurry if you think about the types of *local* adjustments that are typically applied to great taste by professionals… Things like sky/foreground/background adjustments, spot adjustments, dodge and burn, etc.

I feel like it’s a pretty good definition to start with but I’d extend it a little further to encompass the techniques you mentioned and also add things like HDR bracketing, and astrophotographer techniques like dark frames and photo stacking.

Essentially I think as long as the photo is built up from data captured by the camera sensor(s) and no additional data is “added” to the photo, no upscaling, no substitution, no textural or morphological manipulation of objects in the frame, then it’s still a photo even with sophisticated AI processing. But the moment you upscale, airbrush, or morph anything… not a photograph.

That's an arbitrary path-dependent place to put the line. Realistically if you spend infinite time and effort doing the kind of dodge/burn that every photography competition allows then you can make the photo look like anything you want.

One field were ML is making strides apparently, is auto focus systems in mirrorless cameras. Not tgat I would know, so. And tgat is a field I am totally ok with, guessing were the subjects is going, and where to focus on the subject based on user input, is usefull. Shutter, apperture and ISO settings are solved since non-figital cameras got a Program mode. ISO was, obviously, added with DSLRs. It is basically just a stupid control loop running along a curve based on the lense and camera you are using.

> before long we’re going to be at the point where your camera software recognises you’re stood in front of, say, the White House, goes online, grabs stock photography of the White House and merges it into the shot you took.

You realize this is not too different from what the SuperResolution networks are doing, right ?

On one hand, I'm not a photographer, I don't want to be a photographer, and I don't want to spend 30 minutes trying to get everything lined up and everyone looking in the right direction and not blinking or sneezing or whatever. I just want a decent picture that captures people being happy doing a thing.

On the other hand, maybe the world needs more reminders that most people are pretty mundane, hardly able to get more than 2 people looking in the same direction for long enough to snap a shutter, just trying to get through life.

do you really need perfect photos? Does it add anything to the memory?

imho there's value in capturing that someone had a weird expression on their face or blinked.

Yes, for my personal photos. But you have to remember that people aren't taking photos for their own collection, they are taking photos to shine on social media.

Just look on Instagram. How could you compete with all those fit, beautiful and healthy people, if you don't heavily edit your photos.

(To be totally honest, I think that some beauty filters aren't as bad, because most people don't know how to hold their camera and tend to distort their face with closeup lenses.)

Yes, it does add something.

To be vain about it, there's enough drama and messiness in day-to-day life. I don't want to be reminded of that when I look back; I want to see the good parts, that it was all worthwhile.

But, more practically even, photographs capture history. You may not be thinking about it when you're just taking a personal photo, but a fair bit of history does come from personal collections. Especially less significant figures or events where those personal artifacts are the only things that even exist.

Its difficult to gain much useful information when the images are blurry, distorted, or covered in shadow. You may appreciate the funny face, but can a historian or a great-great-grandchild confirm the person in the photo when they are covering half their face in a sneeze?

That said, the pendulum definitely swings in the other direction, too. Despite all the information we gather these days, there may be less meaningful history for the next several decades because you simply can't trust any of it is actually genuine.

We’re there.

https://arstechnica.com/gadgets/2023/03/samsung-says-it-adds...

> I don’t know at what point it stops being a photo any more.

When any form of object recognition is involved.

Wouldn't that preclude using facial recognition for setting focus and exposure though (which has been a thing for years)?

Using facial recognition for focus is ok since the camera still faithfully captures the light entering the lens, but recognizing things in the frame to treat different parts of the image differently in post-processing is a big no-no.

If you don't let it do that, the picture will often be annoyingly dark in the foreground / blown out in the background, because cameras' dynamic range isn't anywhere near your eyes. This is called local tone mapping/dynamic range optimizer.

Sharpening also benefits from object segmentation because you don't want the effects to bleed over into different areas, you get halos that way.

Local tone mapping works just wine without object recognition. It's a dumb algorithm that goes over the image and normalizes pixel values relative to their surroundings to cram the higher dynamic range of the image into the lower dynamic range of the screen. Phone camera apps had it for at least 10 years, and people did it manually with exposure bracketing ever since digital cameras became mainstream.

Yes, but any local area algorithm benefits from segmentation because it automatically becomes "smarter". Upscaling is the most obvious one - the upscaled version of a blue object on a red background is not the average of the two colors - but it applies all over the place.

I don't know of any implementations of this, but an interesting one would be auto white balance. It's typically done as a global slider, but if the image has multiple light sources this doesn't always look good.

And actual straight-up "knowing what's in the scene" AI can help too; people shouldn't look sickly in your photos just because they're under a low-CRI yellow light. You probably want to know what color skin tones actually are.

> Upscaling is the most obvious one

I don't want any upscaling at all in my photos. One pixel on the sensor must correspond to one pixel in the output.

Camera sensors don't have pixels: https://en.wikipedia.org/wiki/Bayer_filter

JPEG files also don't have pixels: https://en.wikipedia.org/wiki/Chroma_subsampling

Then again, OLED displays don't either: https://en.wikipedia.org/wiki/PenTile_matrix_family

I'm aware of all these things but there still is a 1:1 correspondence between camera sensor pixels and JPEG [luminance] pixels. Yes, color information isn't complete, but it's the right balance. Enlarge it and you aren't adding any new information. Shrink it and you're losing information.

Definitely not true unfortunately. In particular it's not true for areas colored red or blue, because the green channel doesn't have the luminance information - that's why I mentioned them.

Lightroom and at least one current phone camera have ML Bayer demosaicing for this reason and it's visibly sharper.

(Note sharpening and enlarging are the same operation.)

Well, if that's the case, there are two options: the photographer git exposure wrong (easy to do in high contrast situations) or you create a HDR photo in post. More often than not so, just changingetering, or the composition by e.g. getting rid of some very bright sky or dark foreground, allows any modern (read: post late 90s) to get exposure right in P-mode. Added benefit of digital photography: you can check your shot on location, including histograms and even live histograms. Takes all of 2 minutes in the field with some practice. Of course, a smartphoe camera won't do any of that.

This is true for dedicated cameras, but smartphone sensors have less dynamic range and so can't naturally get good pictures in a lot of normal situations, especially mixed (indoor+outdoor) lighting.

I would say it's a matter of input vs output. Using AI to identify the best part of the image to act as input is fine (e.g. where to focus, what to base exposure/white balance on, etc). But if you treat that part of the image differently as part of a function's output, e.g. adjusting color differently on a face vs the rest of the image in post, that's where my sense of distaste kicks in. I'd rather that rely only on local color information (e.g. treating a part of the image differently because it's darker is fine).

Even when people were using analog cameras they would edit their negatives before making prints.

It was slower and harder to do than with digital editing, but it was absolutely common-place and normal.

Here's a good overview of some of the old processes and timelines:

https://fixthephoto.com/blog/retouch-tips/history-of-photo-r...

but it was absolutely common-place and normal.

It was slower and harder to do than with digital editing, but it was absolutely common-place and normal.

As someone who grew up in that era, I can say all the family photos have not been edited, nor would the majority of people who owned a camera back then spend the time or money to do so. Yes, commercial photo editing services have always existed, and of course you'd expect things like magazine covers and other prominent, publicly published photography to be heavily edited, but they didn't automatically do it to everyone's photos for free; and that's the huge difference between then and now.

> As someone who grew up in that era, I can say all the family photos have not been edited, nor would the majority of people who owned a camera back then spend the time or money to do so.

Whoever developed your photos would've adjusted the brightness and white balance based on what looked right to them if nothing else. Send the same negatives into two shops and you'd get different-looking pictures back.

Editing -> retouching.

What's the difference?

That the editing/retouching changes small parts of the image and materially alters it to show or leave out stuff that was not in front of the lens and that changing whitebalance and such affects the whole image uniformly.

They used these things: https://125px.com/docs/unsorted/kodak/tg2044_1_02mar99.pdf

It was an automatic process and colours were consistent. I find that reading theories about what might have happened 20 years ago is becoming pretty annoying. I shot some kind of Fujifilm mostly and if it was under/overexposed that was like your problem.

I agree that the average person who owned a camera wouldn't have done this as a matter of course.

Honestly I was thinking more of the "portrait studios", rather than the home-shooter. Though I guess in teh early days of cameras there were a lot of people who did their own development at home, in ad-hoc darkrooms. Maybe they didn't edit so much, but it was within their means, and I'm sure it was more commonly done than we'd suspect.

Retouching was absolutely not a common thing for normal hobbyist camera owners. Most weren‘t even aware this would be realistically possible to do with their photos.

I think the line can be drawn at adding information that wasn't originally available to the camera at that time.

So HDR photography (taking multiple photos milliseconds apart and merging them) is fine, because it's just using information that was available while you were standing there taking the shot. I don't see how this is really any different from a long-exposure shot in fact.

Using a polarizing filter to enhance the sky color and see the fish underwater is fine, because it's just using information available that time (and in fact, selectively ignoring some information, namely light polarized the "wrong" way).

Editing light levels differently in different parts of the frame is OK too because that light is still coming to the camera, you're just processing it differently.

Going online to grab ultra high-res stock photography of the Moon to enhance your shaky-cam photo is not OK, because that's adding information that the camera didn't have available to it (and worse, it's quite possible the downloaded information is inaccurate).

Going back to your White House example, what happens if someone takes a photo of the White House lawn because they want a photo showing the new trees or flowers that were planted or just bloomed. But Samsung downloads some stock photo to "enhance" the photo, and of course this stock photo was taken before the plantings, or at some other time of the year when the flowers weren't blooming. Now this poor sucker has a completely invalid photo showing something that didn't exist when he took his trip to DC to photograph the White House lawn, and probably won't notice until he looks through his vacation photos on a large monitor.

> I don’t know at what point it stops being a photo any more.

Considering that Ansell Adams would spend weeks reworking his negatives to get the perfect print I think that you'll find that horse bolted a very long time ago.

Difference being, he worked in his negatives, he did add some shit to them he found on the internet or some database.

When you are no longer graphing the impacts of photons on a sensor, you are no longer doing photography.

It's just stupid that we gum up relatively simple pipelines of input -> output in favor of these longer and more complicated bullshit... and for what? So unpracticed and uneducated stupi...er, laypeople can feel better about themselves? So companies can cheap out on hardware or optics investments? So some advertising jackass can sell a few more smartphones by manipulating peoples' emotions?

There is so little desire to capture the world as it is. We humans have to insert our bullshit desires and artifice into everything. And winner-takes-all capitalism means the majority market of morons is literally the only market that can be serviced cost-effectively (meaning... at all)

God I hate people so much. So much potential... wasted.

Well I don't think your approach of telling everyone they're stupid is going to sell them a camera…

Who said that I am trying to sell cameras?

I want more investment in better optics and hardware. Not AI hallucinations.

Do you think some random internet comment is going to change anything? Even if it is worded in the best, most persuasive manner possible?

"Hmm, that pdntspa guy on Hacker News is right goshdarnit!", says the Megabux CEO as he jumps out of his chair. "Sheila, get me the engineering director on the line. Tell him to cancel all our AI products!"

Yeah.... not going to happen. Hence, a rant.

Reply on news.ycombinator.com