Is Samsung Faking the AI-Enhanced 'Space Zoom' Photos on Galaxy Smartphones? (appleinsider.com) 95
Samsung's Galaxy smartphones now offer "Space Zoom," writes Apple Insider, a feature augmenting 3x and 10x telephoto cameras with digital zoom "aided by Samsung's AI Super Resolution technology."
But the resulting 100X zoom levels "appear to be more a feat of AI trickery than anything else," they conclude, citing an investigation by a Reddit user: That so-called Space Zoom could potentially allow users to photograph the moon, and many do. However, it may be the case that the level of detail in the moon shots may only be higher due to software shenanigans....
The user tested the effect by downloading a high-resolution image of the moon, then downsized it to a 170 by 170-resolution image, and then applied a gaussian blur to obliterate any final details of its surface. They then showed the low-res blurry moon at full screen on their monitor, walked to the other end of their room, zoomed in on the fake celestial body, and took a photograph. After some processing, an image of the moon was produced by the smartphone, but the surface had considerably more detail for the surface than the doctored source. The user reckons Samsung "is leveraging an AI model to put craters and other details on places which were just a blurry mess."
They go further to stress that while super resolution processing uses multiple images to recover otherwise-lost detail, this seems to be something different. It is proposed that this is a case "where you have a specific AI model trained on a set of moon images, in order to recognize the moon and slap on the moon texture on it."
The Reddit user has now posted an update: I photoshopped one moon next to another (to see if one moon would get the AI treatment, while another would not), and managed to coax the AI to do exactly that.... [O]ne moon got the "AI enhancement", while the other one shows what was actually visible to the sensor — a blurry mess....
It's literally adding in detail that weren't there. It's not deconvolution, it's not sharpening, it's not super resolution, it's not "multiple frames or exposures". It's generating data.
But the resulting 100X zoom levels "appear to be more a feat of AI trickery than anything else," they conclude, citing an investigation by a Reddit user: That so-called Space Zoom could potentially allow users to photograph the moon, and many do. However, it may be the case that the level of detail in the moon shots may only be higher due to software shenanigans....
The user tested the effect by downloading a high-resolution image of the moon, then downsized it to a 170 by 170-resolution image, and then applied a gaussian blur to obliterate any final details of its surface. They then showed the low-res blurry moon at full screen on their monitor, walked to the other end of their room, zoomed in on the fake celestial body, and took a photograph. After some processing, an image of the moon was produced by the smartphone, but the surface had considerably more detail for the surface than the doctored source. The user reckons Samsung "is leveraging an AI model to put craters and other details on places which were just a blurry mess."
They go further to stress that while super resolution processing uses multiple images to recover otherwise-lost detail, this seems to be something different. It is proposed that this is a case "where you have a specific AI model trained on a set of moon images, in order to recognize the moon and slap on the moon texture on it."
The Reddit user has now posted an update: I photoshopped one moon next to another (to see if one moon would get the AI treatment, while another would not), and managed to coax the AI to do exactly that.... [O]ne moon got the "AI enhancement", while the other one shows what was actually visible to the sensor — a blurry mess....
It's literally adding in detail that weren't there. It's not deconvolution, it's not sharpening, it's not super resolution, it's not "multiple frames or exposures". It's generating data.
Re: (Score:1)
I guess the image would still need to have enough similarity with the moon to trigger the algorithm.
The Twitter user who investigated the fake tested that with an image that had one full and a half blurry moons, see "EDIT2 - IMPORTANT" in the twitter post.
He took a shot of the monitor from across the room and the complete moon got Super Zoomed while the half moon stayed blurry.
Come to think of it, there are more ways of testing that. Three days from now, we will have a half moon in the sky. With that, it mi
Re: (Score:2)
Re: (Score:3)
Samsung explains directly they do enhance the moon (Score:2)
Samsung does say they use "detail enhancement techniques" in the last paragraph of their discussion website.
. --> "Finally, we apply detail enhancement techniques that bring out the details of the moon pattern to complete bright and clear moon photos." (translated from Korean)
Source in Korean: https://r1.community.samsung.c... [samsung.com]
Re: (Score:2)
Re:Samsung explains directly they do enhance the m (Score:5, Informative)
What are those enhancements, though? blur reduction/sharpening, brightness, contrast, etc?
Using wget to download a higher resolution picture of the moon, apparently.
Re: (Score:2)
It's all good fun, I don't see a problem with it.
Re: (Score:2)
It's all good fun, I don't see a problem with it.
I bet you would if it was an iPhone.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
(Another person tested it).
It wouldn't be the person on reddit whose post is getting lots of comments from people pointing out potential issues in his methodology, and having issues reproducing his results, would it?
Re: (Score:2)
(Another person tested it).
It wouldn't be the person on reddit whose post is getting lots of comments from people pointing out potential issues in his methodology, and having issues reproducing his results, would it?
I have no clue what you are reading, but I'm not seeing people saying his method is wrong - how many ways can you fuck up drawing a circle and applying some textures? I do see others doing the same thing and getting similar results. Perhaps you can share which link that has "potential issues" and "issues reproducing his results"? https://old.reddit.com/r/Andro... [reddit.com]
Re: Samsung explains directly they do enhance the (Score:2)
What you get with this is a dishonest photograph that is something that looks like the moon but has made up details on it thus it's not really a picture of the moon.
It's like taking a picture of a Toyota and getting a Ford instead.
Re: (Score:2)
They are not 'enhancing details'. They copy-paste the details from a completely different picture of the moon. If something changed about the actual moons appearance then the 'enhanced details' of this algorithm would be oblivious and just reproduce the moon as it was stored in its database.
Re: (Score:2)
Samsung does say they use "detail enhancement techniques" [...]
There's a big different between "detail enhancement" and "Photoshopped", which is effectively what this is.
This is the lunar equivalent of "enhancing" a photo of you spending time with your family by deepfake replacing you with a version of yourself from back when you were 20 pounds lighter and 20 years younger. It isn't "detail enhancement". It's a lie. There's a market for stuff like that, but the fact is that you didn't look 20 pounds lighter and 20 years younger at the moment the image was captured, so
Re: (Score:3)
Wait until this is used to provide criminal evidence, and prosecutors firstly don't reveal the enhancement, and then refuse to reveal the method, and then refuse to reveal the algorithm (which you pretty much can't with AI, that's the point) or that it's AI not multi-data-correlation enhancement.
Re: (Score:2)
which you pretty much can't with AI, that's the point
What? That doesn't make any sense at all. Obscurity is in no way essential and it certainly isn't "the point". No matter the case, the model can always be made available.
Re: (Score:1)
This came up in a recent case. The lawyer prosecuting Kyle Rittenhouse wanted to use a picture that was enhanced using pinch to zoom on an Apple Ipad. The argument against it was that the zoom added details that weren't there. The judge agreed that detail was added by AI and the evidence was not allowed.
Easy test (Score:3)
Isn’t that what all AI upscaling is. Of course, Samsung needs to be forthright as to when AI upscaling starts to get applied during Zoom.
Re:Easy test (Score:4, Insightful)
exactly. guy just could have skipped his whole research plus the text wall just by actually reading the description of the feature ... ... besides just using regular logic, ofc.the whole thing with photographing the blurred picture is just hilarious if not embarrassing, it's useless and proves nothing, that's just a clumsy reproduction of exactly what actually happens when you take a "low" resolution picture of the moon. the "detail" isn't there either. where did he think this magic resolution came from anyway? what do people think "image enhancing" actually means? it's literally inventing pixels based on some heuristics and statistical functions.
well, he surely had some fun, and it's good enough for some slashdot clickbait, apparently.
Re: (Score:3)
AI is being sold as a magic fix it all. Articles of this type are useful in pointing out how much of con that really is.
Re: (Score:3)
Not really. ALL digital image enhancement techniques amount to adding fake data. The only thing this article does is highlight how many people didn't realize that.
Re: (Score:3)
Yes, really. And not just imagery. The public is being sold AI as a magic solution across the board.
Re: (Score:2)
Oh I'm not disputing that side of your statement. I'm disputing the notion that clickbait garbage like this article helps in any way.
Re: (Score:3)
Not all image enhancement. For example, most high end smartphone cameras combine multiple images now. That's not fake, that's just taking more data from the sensor and combining it. Using the data from the sensor directly isn't more "real", because the sensor isn't the same as a human eye and has very specific limitations and distortions. One of the most obvious examples is lens distortion, which is a major issue on phones, especially in macro mode.
Re: (Score:2)
"Using the data from the sensor directly isn't more "real", because the sensor isn't the same as a human eye and has very specific limitations and distortions."
It is more real because it is the direct impression of photons on the sensor. The real outcome of interactions in physical reality. Some enhancements might produce a result subjectively closer to how the human eye sees the subject matter but it isn't doing it by recording actual physics interactions in external physical reality... it is using various
Re: (Score:2)
Not really. ALL digital image enhancement techniques amount to adding fake data.
Oh really? So if i make an image brighter so i can see better then i'm adding data?
If i play back an audio track and i increase the volume to hear a soft part better, then i'm adding fake data?
Re: (Score:2)
That he happens to be right about the broader point, that the article and other AI articles are selling us a load of bull, is likely more accident than insight. Even a blind squirrel gets a nut once in a while.
Interesting take! :)
Re: (Score:2)
"Oh really? So if i make an image brighter so i can see better then i'm adding data?"
"If i play back an audio track and i increase the volume to hear a soft part better, then i'm adding fake data?"
The depends on how you did it. If you adjusted your monitor or turned your volume knob then possibly not depending on the type and quality of amplifier in the your sound system but those aren't digital enhancements, nor even changes to the recorded material. Those are analog.adjustments to the output. If you made
Re: (Score:2)
Those are analog.adjustments to the output.
It wouldn't change things whether i'd use an analog or a digital volume knob. Neither adds 'fake' data. A volume operation doesn't change the informational content of the signal. It only makes the information already contained in the signal easier to perceive.
If you made those adjustments digitally then yes you created fake data...
Nothing (well, in practice very little) about the informational content changes due to a scaling (volume change), whether analog or digital.
And in fact, I can guarantee you the analog version is faker than the digital one. Amplifiers don't actually pas
Re: (Score:2)
"I chose volume because it is a linear operation that both doesn't change the signal information wise and is used as an enhancement of audio."
Yes, you chose linear amplification because the entire point of linear amplifications is to change one property while leaving the others unchanged, the new value of the changed property... fake data. ;) It is cherry picking because every other enhancement will then be discriminating and selectively modifying properties for enhancement which is more obviously fake data
Re: (Score:2)
Seems like it's doing what it's supposed to? You wanted a good picture of the moon, it made it happen.
Re: (Score:3)
No, that's not what it's supposed to do. You didn't want a good picture of the moon. You wanted your picture of the moon to be good.
Re:Easy test (Score:4, Insightful)
it's useless and proves nothing,
Au contraire. It proves that samsungs marketing claims for 100x zoom are pure bullshit.
where did he think this magic resolution came from anyway?
From samsung claiming the camera has 100x zoom, maybe?
what do people think "image enhancing" actually means?
It can mean a lot of things. Most often it means manipulating the data of the picture so that more information can be gained from the picture by the human eye. Which is not true for the AI upscaling as the details are not representing the original captured data anymore but are mere suggestions of what could have been there. So samsung calling it 'enhancement' is really a degradation of the meaning of 'enhancing'. It is 'enhancement' but only very superficially and without actually using the information in the picture data. It is a vanity-type of enhancement and the result is further from the truth than the original.
Re: (Score:3)
Bingo. This is a non-story. I'm not sure what he thought AI enhancement or upscaling was... ALL upscaling is some kind of algorithm inserting fake data. Where did they think the additional detail came from, magic?
Re: (Score:2)
Very few people understand what an OTF is. All interpolation is making shit up. The AI stuff just makes a guess based on experience rather than guessing the pixel to the left, or splitting the difference, etc.
The pictures in the article aren't really that impressive. That blurred moon is still clearly recognizable.
Re: (Score:3)
In this case they seem to just be using AI to recognize the Moon. Then they replace it with a higher resolution version, and apply suitable shading for whatever phase it's in.
It's a good idea and could be used elsewhere. Say you are at some landmark and want to take a selfie. The phone could remove other tourists by substituting real data from a reference image of the landmark, or help with zooming etc.
"Writes Apple Insider" (Score:2)
Hm, a report by "Apple Insider." Gee, I wonder if an Apple fansite might be biased against an Apple competitor, especially given that Apple's camera has lagged behind Android for years now. The iPhone's camera is still one of the worst smartphone cameras and an absolute embarrassment for a "flagship" phone. I guess not everyone can be as "innovative" as Apple with their rounded corners and brand new "pro-only" always on display, and instead focus on things like a camera that produces quality photos and an O
Re: (Score:2, Informative)
The iPhone's camera is still one of the worst smartphone cameras and an absolute embarrassment for a "flagship" phone.
https://www.pcmag.com/picks/th... [pcmag.com]
https://www.cnet.com/tech/mobi... [cnet.com]
https://www.trustedreviews.com... [trustedreviews.com]
You can argue all you want that it's not the best (although it is certainly in the running) but suggesting it has "horrible color" and "over-smoothed" etc etc is silly. And "Android phones" overall are nowhere near as good as Apple, maybe you meant compared to the upper-end flagships from Samsung and Google? But are you really suggesting that _all_ android phones have better cameras than the iPhon
Re: (Score:2)
"But are you really suggesting that _all_ android phones have better cameras than the iPhone? Surely not."
You might have a better case if the Samsung and Google phones were the elite or something but the reality is that they are the only android phones which are relevant in the market and virtually all android phones ARE samsung or google phones.
Re: (Score:2)
The iPhone's camera is still one of the worst smartphone cameras and an absolute embarrassment for a "flagship" phone.
/quote
working as intended. iphone is king of phone cameras by a long shot for what most people actually expect from phone cameras: foolproof delivery of flashy, captivating images at any conceivable crazy angle, optimized for social media streams. they do some impressive work with light, color and focus for that. nearest contender is probably the google pixel and still at some considerable margin. ofc if you're a professional photographer neither is for you.
Re:"Writes Apple Insider" (Score:4, Informative)
My dSLR's, and even brand new releases of modern big cameras, still aren't doing computational photography.
> optimized for social media streams
That's definitely another factor. Processing a raw image to be usable is too much work. iPhone fakes a good enough take right out of the gate and the iOS editor has enough sliders to tweak it to a final form. Is it perfect? Hell no. But it saves me dozens of hours per week in editing.
Re: (Score:2)
As a pro photographer I've found myself using iPhones more and more.
My dSLR's, and even brand new releases of modern big cameras, still aren't doing computational photography.
Yes of course that's what RAW format is for
Re: (Score:2)
As a pro photographer I've found myself using iPhones more and more.
My dSLR's, and even brand new releases of modern big cameras, still aren't doing computational photography.
I've found myself realising that phones do not give me any options over "computational photography" and I leave that stuff to the PC where it belongs after taking a RAW file from my DSLR. The PC where many applications are available to do this for you, e.g. Topaz Denoise AI (which I can highly recommend despite being a sceptic at first)
I don't want to gatekeep, but it sounds like you as a pro photographer are very much acting like a happy snap family man of the years of old, point a camera (however competen
Re: (Score:2)
Re: (Score:3)
I've been wondering who DSLRs don't do computational photography. I heard some Sony models are starting to get there, but I'd be interested in a DSLR that can produce higher resolution images than a phone but with the same ease of use - point and shoot hand-held, sorts everything out for you.
Moonify (Score:2)
If Samsung is "cheating", it's really not much different than when streaming services identify a song playing from a tinny distorted speaker and let you listen to the full original version. It's not as if any of us are able to travel to other planets and photograph a different moon. This would be big news if people were on the star ship Enterprise and all their photos of alien moons turned out to just be AI-enhanced versions of Earth, but since we're all presently stuck on Earth...
For that matter, it's al
Re: (Score:2)
thank god that's basically what samsung is doing here. Every idiot points their shitty-ass phone at the moon and tries to take a picture. It is purely logical to use AI to inference that and give the customer what they want, just copy paste that high res NASA moon photo in the customer image, and call it a day. Their idiot brains will get the
Re:Moonify (Score:5, Interesting)
Usually landmarks don't change but when they do, people want to capture them.
Re: (Score:2)
Fair enough but it isn't a real 'gotcha' RE Samsung. ALL image enhancement filters are injecting fake data whether it is a sharpen, a blur, a color correction, an upscale, etc it is FAKE. The only big difference is instead of working off nearest pixels and human hand-coded mathematical patterns and tricks they are using a giant database stuff and making decisions on what has the highest probability of being the right answer based on that training data. That is all "AI" really is.
If you don't want fake data
Re: (Score:2)
Not false. And nothing of the sort has been explained by anyone... or even claimed to have been. But even if had your comment would entirely redundant and pointless.
Re: (Score:1)
Re: (Score:3)
It's the moon, it looks the same in every photo.
Unless, perhaps, you are in the southern hemisphere. There the moon appears upside-down (relatively speaking) and the phases happen in the opposite direction.
Though I could assume that Samsung is smart enough to recognize this. On the other hand, that would involve programming, and we all know developers always write perfect programs that account for every edge case...
Well duh! (Score:3)
Did he really think it actually recovered data that no longer existed in the image? Of course its AI sleight of hand! Either it loads up what it thinks is the original image and slaps that on top, or it simple does some intelligent interpolation - ie makes it up. Theres no 3rd option.
Re: (Score:2)
>>intelligent interpolation - ie makes it up
If the Samsung AI has added details that didn't exist, I think and is interested to see test on human faces to prove this.
By taking something like a passport photo on a wall from afar, where the Space Zoom made details on assumptions, there is a very high chance the resultant photo would be an AI generated human face of high resolution.....that also NOT look like the original person on photo at all.
Re: Well duh! (Score:2)
the reason it "clicks in" on the moon specifically is probably because the models have been trained to NOT do that to faces, because Samsung knows that would be a PR disaster.
my guess is there's a layer to do a whitelisting of acceptable substitutions (=landmarks), output of which is passed on to a standard AI upscaler, which is also restricted to adding from acceptable details.
How dare they! (Score:3)
This is outrageous, clearly they are using AI based image regeneration and not a real optical lenses. They are cheating. The devious scum. I mean in all their marketing they stated ... well... stated ... erm ...
100X Space Zoom includes 10x Optical Zoom and 100x digital zoom with AI Super Resolution technology.
Wait so... We caught Samsung doing exactly what they say they are doing on the product page? I guess this is news given how dishonest they were in the past about their camera capabilities, but the headline should really read: "Samsung does a first by producing a product that works as advertised."
Re: (Score:3)
The point is that they use the moon shot a LOT in their advertising, and it is not representative of what the phone is capable of doing. For example, if you tried to take a pic of some similarly-distant (or small on the sensor) object, you won't get the results the moon pics appear to promise. Essentially, Samsung has special-cased the moon by including a very detailed picture of it in the algorithm – they didn't include pictures of, say, the Eiffel Tower or the Statue of Liberty, and a picture taken
Re: (Score:2)
Re: (Score:2)
Samsung has special-cased the moon by including a very detailed picture of it in the algorithm
Not really. Samsung has just shown off the best example of it's system, something that is done by literally every company in every field (not just phones or technology) for the entire history of humanity. Just because the moon gives the best results doesn't mean that the system isn't a well trained AI enhancement based system like they advertise on their page, and AI enhancement of blurry pictures does absolute frigging magic in many cases (I personally use AI enhancement tools as well, just on the PC with
Re:How dare they! (Score:4, Interesting)
they didn't include pictures of, say, the Eiffel Tower or the Statue of Liberty, and a picture taken from a great distance of them is still going to be a smear of interpolated goop instead of the crystal-clear quality promised in the ads.
Just imagine if they did. And then imagine the shit storm if some terrorists take down the Eiffel Tower, but it still continues showing up in Paris skyline photos taken by tourists with their Samsung afterwards.
Re: (Score:2)
Kind of a silly way to frame the issue.
If it were truth in advertising it would be 10x Optical Zoom and 100x digital zoom with AI Super Resolution that only works on a couple of pre-defined objects.
It wont work if you try to take a picture of an aircraft, or spy balloon, or other aerial phenomenon or basically ANYTHING that isn't the moon or a couple of well known star constellations.
Re: (Score:2)
If it were truth in advertising it would be 10x Optical Zoom and 100x digital zoom with AI Super Resolution that only works on a couple of pre-defined objects.
No it doesn't work only on pre-defined objects. AI Super Resolution works on all images. How well it may work is highly dependent on the model, and precisely 100% of all companies in history ever have advertised their products using examples of when they work best. The moon just works best, point the camera at anything else and the algorithm still does it's thing.
The biggest problem here is the lack of a user selectable and tweakable model which you would get with commercial PC software.
In short, yes. (Score:2)
I think Samsung's had the 100x space zoom since the S20 Ultra? I had the S21 Ultra and now I have the S23 Ultra. It's very cool, I love the zoom (its useable up to 30x, which is 3x digital zoom on the 10x optical zoom camera) but starts to get very pixelated above that. 100x is not very useable.
But I feel it's been pretty obvious to everyone that Samsung is basically faking the moon photo by overlaying it.
Not sure why this is suddenly news now?
Re: In short, yes. (Score:1)
Would have been a hit product if they used ChatGPT (Score:2)
ChatpGPT is the new blockchain.
Test it with a Death Star image (Score:2)
Re: (Score:2)
How is this going to work with people? (Score:3)
Are we going to take a zoomed in picture of a person with long dark hair, and have the AI paste in the face of one of the Kardashians?
Is the moon a secret ? (Score:2)
Yes? (Score:3)
That's what Ai enhancement is. How do people think this stuff works?
Moon (Score:2)
Surprise, that's kind of how these things work. With millions of sample pictures, it can find segments that, when blurred, approximate the blurred thing, and hence "unblur" it.
If it can recognize the object, all the better.
Never mind the moon... (Score:2)
Dishonest photographs (Score:2)
"It's literally adding in detail that weren't there. It's not deconvolution, it's not sharpening, it's not super resolution, it's not "multiple frames or exposures". It's generating data"
I'd rather have a "blurry mess" than a picture that is an outright lie.
I take pictures to have a record of something real. If I want fantasy art, I can find it anywhere or create my own.
If they are selling this whold deal as an "enhancer" rather than "making up details to make the picture look good"
A "photograph" is a record of light (Score:2)
Re: (Score:2)
Photos of family members start looking like sit-com actors.
Re: (Score:2)
Re: (Score:2)
You need to understand that all digital cameras lie. Even your raw files directly out of the camera have some processing on them. That's why all the camera vendors talk about their new image processing chips more than the sensor.
Re: (Score:2)
Same thing happened with Huawei P30 Pro in 2019 (Score:1)
Deckarded (Score:2)
Enhance 224 to 176. Enhance, stop. Move in, stop.
AI literally is trickery. (Score:2)
The Moon never looks the same twice (Score:2)
We're generally taught that the same face of the Moon always faces toward Earth, but that's only an approximation. There are a few effects that change the appearance of the Moon, so it'd be interesting to see how the AI reacts to those.
Obviously the phase affects how the Moon looks. Another is called "libration", due to the fact that the Moon's orbital speed varies, while it's rotational speed remains fairly constant. And the fact that the Moon's orbit is inclined 5deg. Parallax is the fact that two ob
try a different moon (Score:2)