[drewprops]

Tonight I learned that Samsung has been caught in "Moongate" (or whatever it will be called).

The latest Samsungs are apparently delivering INCREDIBLE photos of the moon.

Samsung has denied that it is doing anything in software.

A redditor created a low-rez image of the moon, put it on their monitor, went across the room and zoomed in on it.

The image the phone created had fabulous detail.

Turns out they're using AI to cheat on creating those images.

...

[PB PM]

Hard to attack Samsung when, Google (been doing it for years on Pixel phones), Apple and all the other phone makers are doing the same thing to make smartphone pictures look better with newer models.

[drewprops]

I look forward to more about this coming out on all the platforms - articles about "which smartphone screws with your photos the least?".


...

[chucker]

Originally Posted by PB PM Hard to attack Samsung when, Google (been doing it for years on Pixel phones), Apple and all the other phone makers are doing the same thing to make smartphone pictures look better with newer models.

Apple does not fantasize image elements that aren’t there.

[drewprops]

I do agree that it seems that a line has been crossed. Clearly, Samsung isn't just juicing color balance or softening faces - they are adding data that was never in the image to begin with.

If I'm a defense lawyer, I'm going to call into question any evidence submitted by a Samsung smartphone of this generation, and will attempt to indicate that ANY images taken by smartphones are unreliable, as many Google phones allow you to remove people and backgrounds.


...

[PB PM]

Originally Posted by chucker Apple does not fantasize image elements that aren’t there.

All digital cameras do, and always have, always will. Not with AI, but yes they do.

As for Apple, they do all kinds of stuff with machine learning to make pictures look sharper and less noisy than a sensor the size of the ones in smartphones at capable of doing. They also automatically apply filters, depending on the selected photo mode. So let’s not pretend they aren’t doing anything. They haven’t gone as far as Samsung and Google, but they likely will.

[drewprops]

Perhaps they will not?

Tim Apple has made privacy and security a selling point with the company's products.

This would be an opportunity to make a statement about not delving too far with sweetening images.

What do you think about that?


...

[PB PM]

Apple has to stay competitive with the cameras or they will lose customers. Simple as that. People who aren’t into tech don’t care if the image is AI corrected, they want the best photos they can get.

[chucker]

Originally Posted by PB PM All digital cameras do, and always have, always will. Not with AI, but yes they do. As for Apple, they do all kinds of stuff with machine learning to make pictures look sharper and less noisy than a sensor the size of the ones in smartphones at capable of doing. They also automatically apply filters, depending on the selected photo mode. So let’s not pretend they aren’t doing anything.

Which is completely different than “it looks like you’re trying to photograph the moon. Here, I’ll substituted a sharp picture of the moon in its place.”

Originally Posted by PB PM They haven’t gone as far as Samsung and Google, but they likely will.

Google does nothing of this sort.

[PB PM]

Originally Posted by chucker Google does nothing of this sort.

Google has been using AI to improve low light images on Pixel phones for years, and we aren’t just talking about noise reduction.

Look we can argue about the finer points of the type image manipulation in photos coming out of different brands, but at the end of the day they are all doing it.

[chucker]

Originally Posted by PB PM Google has been using AI to improve low light images on Pixel phones for years, and we aren’t just talking about noise reduction.

Which, again, is very different. They’re not adding objects and textures that aren’t there.

Originally Posted by PB PM Look we can argue about the finer points of the type image manipulation in photos coming out of different brands, but at the end of the day they are all doing it.

Applying ML to the image pipeline? Yes. But not at this level.

It’s a story because it goes much further than others, and arguably too far. On top of that, Samsung seems to be lying about it. They call it “Space Zoom”, implying it happens thanks to the zoom lens, and they claim “no image overlaying or texture effects are applied”, which this suggests is flat-out false.

(You could argue, “ah, but there’s no literal code path that inserts the moon texture; it’s just that the AI has been trained to think it ought to be inserted”, and to that I say it’s a distinction without a difference. Apple’s Night Mode and Deep Fusion or Google’s Night Light don’t insert additional patrons when you’re taking an indoor photo at a bar. Samsung does, however, appear to be inserting such information.)

[Bryson]

In my opinion the fact that they do this is a non-issue, and actually a pretty interesting idea.

The fact that they lie about doing it, however, is not acceptable.

[Mac+]

Dare I say, not new either for them.

[PB PM]

Originally Posted by chucker Which, again, is very different. They’re not adding objects and textures that aren’t there.

My understanding is that Google is using an image database to improve night photos. How is that not the same thing?

[PB PM]

Originally Posted by Bryson In my opinion the fact that they do this is a non-issue, and actually a pretty interesting idea. The fact that they lie about doing it, however, is not acceptable.

It’s not even new. This type of AI image enhancement has been available for desktop users of several photography software packages for years now. They can take low resolution images like the one in the link article and make them as good or better than what Samsung is doing.

[chucker]

Originally Posted by PB PM My understanding is that Google is using an image database to improve night photos. How is that not the same thing?

Like I said. If Google's algorithm caused artifacts from that image database to appear, I'd consider it the same thing. For example, if a scene from a night club shows more alcoholic bottles and more patrons than were actually there, because the average image from the database has more.

[drewprops]

Insertion of nonexistent data is different.

You cannot equivocate.

...

[PB PM]

Originally Posted by chucker Like I said. If Google's algorithm caused artifacts from that image database to appear, I'd consider it the same thing. For example, if a scene from a night club shows more alcoholic bottles and more patrons than were actually there, because the average image from the database has more.

That's exactly what it does though, it adds detail from the database that would be missing in a normal night photo from a smartphone. This Samsung AI isn't adding things to the moon that don't exist, it's just adding data from a database to someones photo. It's completely the same thing.

[drewprops]

I am admittedly NOT someone who tracks what has been happening in smartphone photography space, and am not super familiar with the adjustments these gizmos are making on the fly on my behalf.

Can we use some concrete comparisons?

The nightclub and alcohol bottles example is helpful.

Is MY iPhone inserting non-existent imagery into my photos?

I really want to know about it, if it is.




...

[PB PM]

Originally Posted by drewprops I am admittedly NOT someone who tracks what has been happening in smartphone photography space, and am not super familiar with the adjustments these gizmos are making on the fly on my behalf. Can we use some concrete comparisons? The nightclub and alcohol bottles example is helpful. Is MY iPhone inserting non-existent imagery into my photos? I really want to know about it, if it is. ...

As for as I know Apple is only using machine learning, not AI correction from an image database. Machine learning is basic form of AI, allowing programmers to improve algorithms. Over time the Apple system will improve images based on what Apple says it can. The AI kind of learns and implements improvements itself. At least that's my rather surface level understand of the difference from my friend who is into this stuff.

So is Apple doing things to enhance the images your iPhone's camera takes? Yes. How much? Apple decides and changes it over time. Are the using an image database to improve your photos? I'd say there is a very good chance that they are using some kind of image database. Every camera maker is doing that, from smartphones to dedicated cameras like your Fuji. Back to Apple though. How much are they doing to your photo, and will you'd notice? Unlikely you'd notice unless you could get your hands on a RAW file from the camera. If you use the portrait mode you likely will see skin smoothing and such things going on.

[drewprops]

Thank you, that is a helpful place to stage our conversation from.

Image replacement is probably an issue now, because of Moongate (or whatever).

Might there be a chance that we are offered an easy to pick "reality mode" option, in the future?

Like I said, I don't see how photographic evidence will be permissible 10 years from now.

The world gone mad.

...

[PB PM]

Photos have been manipulated since they were first made, first in the darkroom, now by AI, all that has changed is the method.

There are ways to find out if an image has been manipulated, but it can take a sharp eye to notice when fully zoomed in. Right now AI kind sucks with people, they often mess up ears, or fingers. I’m sure that will change with time though. To me AI voice spoofing is far more dangerous on its own, but when we combine fake photos, videos and voices like they are now, it’s clear we are headed to a scary time in history. The

[Matsu]

I think there's an important distinction between using learning to enhance or optimize recorded data versus using data recorded from other sources to supplant data that could not be (and was not) recorded by the system in real time. Depending on implementation, the former may (positively or negatively) alter or effect the representative value/utility of the image for documentary and forensic purposes. The latter destroys it.

[PB PM]

Given that all digital cameras create pixels in the gaps between diodes in the sensor, even in the so called gapless sensors, in essence they are all doing it.

The question, which are article fails to prove, is does the AI create something or enhance with the database? As noted before there are several desktop apps (such as software from Topaz Labs) that can turn a blurry photo like the one in this story and up the resolution and sharpen. No reason that Samsung couldn’t be doing the same, rather than the accused outright replacement theory.

[Matsu]

Only in a very grossly analogous sense can we claim that de-mosaicing is the same as introducing external data into the scene. It's not the same thing at all. The limits of what can be resolved for a given sensor and optical system are understood. When we know where the system limits are, we can use calculations to help us see just a little better, to enhance or visualize the recorded data/"graph". In so doing we have a good understanding visual phenomena recorded from the natural world vs artifacts introduced by some limit of the system and/or overprocessing of it's recorded data. Think of moire patterns, LoCAs, shutter speed effects like wobble, and others.

There's a world of difference between doing best guesses (mathematical computations) based on what we caught in the frame, and adding supplemental external data to the frame.

[PB PM]

Originally Posted by Matsu There's a world of difference between doing best guesses (mathematical computations) based on what we caught in the frame, and adding supplemental external data to the frame.

Thing is they have all, yes even Canon, Nikon, Fuji, and Sony, been doing this for years, nobody said boo about it. All digital camera have had ever growing image databases in them. First they were used for improving focus performance, then noise reduction, and then to improve dynamic range. The skin smoothing, AI texture sharpening, improving and yes enhancing are simply the next step, difference is the likes of Apple, Google and Samsung are taking the lead now.

Are there ethical issues? Of course. Reality is, the AI genie is out of the bottle and it isn’t going back in. I mean Microsoft just laid off their AI ethics team.

[Matsu]

That's not true. There's no evidence of any external data being added on to the in-camera RAW capture from system cameras of the major manufacturers. There are geometric corrections, vignette, chromatic aberration, noise/smoothing being applied, but these are not "find and replace" routines. They are a mathematical look at the data and what can be enhanced using a very complex understanding of the recording and both the sensor and optical system used to make it. They're not inventing pixels from some external data set and replacing what the camera sees.

Software on the other hand...

[Kickaha]

From my perspective, it's the difference between scanning a typed document, and having the OCR fix the occasional blemish or recognition error because "Sorry, but çat is not an English word, so I'm turning that into cat", versus having the OCR change the meaning of the sentences by deleting or including select words without warning, or adding entirely new chunks of text that were not in the original document.

Both are altered, but one is intended to clarify what is there, while one is intended to fundamentally alter what the subsequent reader believes to have been in the original.

These are in no way equivalent or comparable.

Same with the photos. To claim that removing sensor noise and altering what the photo contains in frame, is merely a matter of degree, is at odds with reality.

[Brad]

Originally Posted by Kickaha From my perspective, it's the difference between scanning a typed document, and having the OCR fix the occasional blemish or recognition error because "Sorry, but çat is not an English word, so I'm turning that into cat", versus having the OCR change the meaning of the sentences by deleting or including select words without warning, or adding entirely new chunks of text that were not in the original document.

Don't feel left out! Pretty soon ChatGPT/GPT-4 will be seamlessly embedded into your word processor of choice, and it'll do exactly this for everyone all the time.

[PB PM]

Originally Posted by Matsu That's not true. There's no evidence of any external data being added on to the in-camera RAW capture from system cameras of the major manufacturers.

Never said they did, I said they have on board image databases. I've been talking about the use of image databases the entire time. Nor did I say they were using AI. Read your DSLR or mirrorless cameras manual, it’s covered that they have on board image data sets, going back to 1999. How many images and what they are used for has changed over time.

Did you think those camera makers were improving image quality they way they have just from sensor improvements? They are running a ton of processing, yes even on RAW files.