The year is 2036, when dialed up to maximum space saving, the latest generation of smart phones now turn camera pictures to a textual description and later when you want to view them, re-generate them on the fly using a text2image model. Not that "saving space" really means something, the phones refuse to unlock if they don't have an internet connection and store all your data on multiple cloud servers anyway (one for each five eye country).
That one monarch is not a person or organization of people. It's an aggregation of all the loosely connected ones, competing on some things, coordinated on others, but you, the end user are never the beneficiary. That, my old school human friend, is the one commandment no one challenges. With all those players locked into a Nash equilibrium, no one can do anything about this, while all humans complain that things are not the way they should be.
There was a post on HN a few months ago showing the stages of post-processing on a photo of a Christmas tree in a room - the original photo was unrecognisable.
I would find the link, but I’m balancing my phone on my knee while trying to eat dinner and pretend I’m not playing with my phone while eating dinner…
With the rise of AI-generated content I would actually love a camera with no post-processing by default and a watermarking system that proves the picture is authentic. I can see Apple going that way eventually.
Every dedicated camera is like this, more so if you shoot in raw.
There's also a lot of confusion around the processing here. The iPhone is mostly just taking multiple pictures and stacking them on top of each other to make up for the poor performance of the small hardware, denoising, and then applying a mask to subjects to lift shadows and adjust skin tones. This is the same kind of edit a photographer would do in Lightroom. It's not straight up AI generating content in the image.
Sadly Android phones are willing to pull cheap tricks to show better numbers or look better in some comparative youtube video so they pull this stuff.
Leica and Sony offer this as “Content Credentials” and anti-forgery digital signatures done on camera. It’s also been standard for many years for digital cameras used in criminal evidence procedures, Olympus tough series IIRC.
Me I just switched back to film. It’s about a buck a picture but I like the results I get, I’d rather have 30 photos than 300 to look through anyway.
This is a feature a lot of companies are looking in to right now. It's less about reducing processing and more about proving something is real and not AI generated. The iphone processing doesn't straight up generate things that aren't real.
It's also something I've been looking into. I've completely broken Google's implementation and I can sign any file as if it were real, more details will be public in the coming weeks.
Signatures are nice, but the contents of the image still matter. Any signs of manipulation should still be treated with suspicion, even if they're "legitimate" edits. The best way to avoid such signs is to have the bare minimum processing.
This is also my opinion. However, I still think it's worth raising the bar for plausible fakes. Minimising processing is another way to raise that bar (or at least, lower the floor). It's also not something you can do by default, because consumers demand image processing.
Telling if a photo or a video is "true" is the same as telling if the written text is "true". You can not tell it by analyzing characters / words in a text, or pixels in a video.
You should analyze the subject who is providing that text / video and decide for yourself, whether you trust that subject or not.
E.g. "Elon Musk is 10 feet tall" - are you reading it on an anonymous X account or on a Business Insider social account? An attached photo of Elon Musk being 10 feet tall does not make a difference.
Well, they do their best, they don't do any userspace post-processing. They can't really help whatever the firmware in the camera and/or the camera driver are already doing.
It's not really possible to have zero processing of an image. At some point you have to interpret the sensor data and turn it in to a JPEG. The line people decide an image is "processed" is pretty subjective though. It's somewhere between the camera deciding what the white balance / exposure is, and AI replacing the sun with a moon.
The line I draw is at semantic editing. Manipulating images based on abstract geometric or statistical features like edges or color histograms is acceptable. Attempting to guess what those features mean is not. Traditional sharpening, color balance correction, focus stacking, lens distortion correction, etc. is fine. Red eye removal, skin tone correction, traditional dodging and burning, replacing a detected moon with a higher resolution photo, etc. is not. Unlabeled semantic editing is dishonest. AI image processing does not distinguish between geometric and semantic features, so it's always unacceptable, even for simple tasks like sharpening.
The one exception I make is cropping, because photographers have always had the ability to choose where to point the camera. Cropping based on meaning is not dishonest because it's inherent in the process of photography.
Any file format or data structure you can display on the screen is processed. The whole point is the debate around phone processing is not as clear cut as photos being processed vs unprocessed.
I am wondering if the AI enhancement should be opt-in rather than the default and saving the original as well. Even if you want AI enhancement, future models are likely to be better.
Interestingly, one of their AI filters is actually called "Super Moon":
4.14.3 Camera Image Optimization
The camera image optimization AI algorithm is mainly used in the following camera features: Super Moon, AI Camera, Document Mode, Front Portrait Mode with Background Blurring, and Beautify.
I still struggle with the idea that the chinese olympics had both CGI fireworks and real firework footage mixed together.
I cant fathom how this was a "if they know it doesnt matter at all" situation.
A complete acceptance of real / fake, who cares its about stuff looking great.
I mean I can be told it but I cant understand why you would waste millions on real works just to have fake ones, why not just go all fake? If its a half way house then does no one care what is real?
IE on linkedin somany AI videos with thousands of comments and likes and I become lost on ... do most poeple just not care if its real or not? I dont think they do at all. Am I mad to care?
Its very easy to say well everyone uses filters and they are not real. And its true! But somewhere and I have no idea where, there seems tobe a line for me where I say no, thats not cricket, you cant just replace part of an image with a template of something else. I have no idea tho any thoughts appreciated.
> I mean I can be told it but I cant understand why you would waste millions on real works just to have fake ones, why not just go all fake? If its a half way house then does no one care what is real?
I work with VFX and try to always have something real in the shot. Explosions and the like are easier to expand but quite difficult to just add to empty shot. Fireworks are a bit easier but there is interplay with the environment which is hard to fake.
All these films where they announce ”we did it for real with practical effects” usually mean that it’s a mix and match between practical and computer generated effects. It can look better than either on their own.
Also I guess they had audience there so you need something for them as well.
Perhaps they needed some real fireworks for those there in person, and to be able to claim that there really was a fireworks show, but wanted to amp it up for those viewing the event virtually (most of the world).
From what I have seen, the attitude towards AI in China is entirely different. The population largely believes it's a positive technology that will help improve their lives. While in America there is a seething hatred towards AI where people believe the benefits will all go to the billionaires while they are put out of a job.
If you look at the way Chinese AI companies act compared to American ones, it's easy to see how people feel this way.
The crucial difference: in China, if a billionare might become more powerful than the government, they get disappeared and re-educated: https://www.bbc.com/news/technology-56448688
Jack Ma is as scummy as the American tech elite, The whole 996 work schedule was from his companies. It's honestly good for society that people like him get reined in.
And it's not like they sent him to some north korean death camp. He's still one of the richest people in China, he just failed to deregulate the finance industry.
i would not underestimate the tomfoolery Xiaomi is doing to give the impression their phone can take a zoomed in photo of the moon. It's pretty tasteless in my opinion
I got some amazing photos through a set of eclipse viewing glasses on my Vivo x300 pro. If I'd been a bit more organized and set up a stand and taped the glasses onto the phone I think I could have taken a video of the whole thing in incredible quality. But I just held the glasses in front of the lens and it worked.
What I found though was that I had to use the pro mode (which I use for basically everything anyway, to avoid exactly the kind of issues this article is talking about).
I did try on normal photo mode and I think what was happening was that as soon as I covered one lens with the glasses it switched to another.
I don't think it's intentional. This behavior is a straightforward consequence of training something on pairs of "real" and "degraded" images. It learns what images look like and does the best it can. Philosophically, replacing the moon with a better moon is no different than any other computational photography technique that uses world knowledge, like reducing red eye from flash by detecting faces, or enhancing blurry text using knowledge of the alphabet.
The problem is philosophical - people are upset by certain types of enhancement error, but not others, and they can't even articulate the difference to themselves, let alone to an AI.
> people are upset by certain types of enhancement error, but not others, and they can't even articulate the difference to themselves
People are upset by enhancement errors that produce an image that is meaningfully different from the "true image". Replacing the sun with the moon or replacing the letters with different letters (e.g. https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres... ).
The philosophical problem here is that you can't get at the true image, you can only imagine ideas like "what the human originally taking the photo saw with their naked eye".
What constitutes a "meaningful" difference? If you replace flash-induced red eyes with a more plausible (yet inaccurate) eye color, people are perfectly happy with that, even though the camera indeed saw red photons. They're not happy if you run the exact same algorithm on a photo of someone with genuinely red eyes. It's not always about the "true photons" - the cause of the red eye is relevant, and so is the intent of taking the photo.
"what the human originally taking the photo saw with their naked eye" is a tempting standard, but a misleading one - there might not be a human eye near the camera at all, and at any rate there certainly isn't one co-located with the focal point. And shall we have cameras that render only a few degrees of high resolution color, and have a large-ish blind spot? It's what the human eye sees.
In the future, it might be extremely hard to find a camera which records exactly what is in front of it :D
The year is 2036, when dialed up to maximum space saving, the latest generation of smart phones now turn camera pictures to a textual description and later when you want to view them, re-generate them on the fly using a text2image model. Not that "saving space" really means something, the phones refuse to unlock if they don't have an internet connection and store all your data on multiple cloud servers anyway (one for each five eye country).
It's already 14 eyes, and it'll be 195 eyes by then
194. That one country will not be invited.
It's one Monarch, but many eyes. All information flows to the Crown.
That one monarch is not a person or organization of people. It's an aggregation of all the loosely connected ones, competing on some things, coordinated on others, but you, the end user are never the beneficiary. That, my old school human friend, is the one commandment no one challenges. With all those players locked into a Nash equilibrium, no one can do anything about this, while all humans complain that things are not the way they should be.
A lot of smartphones support raw format.
It's fascinating to see how much post-processing gets done.
There was a post on HN a few months ago showing the stages of post-processing on a photo of a Christmas tree in a room - the original photo was unrecognisable.
I would find the link, but I’m balancing my phone on my knee while trying to eat dinner and pretend I’m not playing with my phone while eating dinner…
It's this https://news.ycombinator.com/item?id=46415225
Often the "RAW" output is still DSP'd, just less (e.g. only debayering + noise reduction).
With the rise of AI-generated content I would actually love a camera with no post-processing by default and a watermarking system that proves the picture is authentic. I can see Apple going that way eventually.
Every dedicated camera is like this, more so if you shoot in raw.
There's also a lot of confusion around the processing here. The iPhone is mostly just taking multiple pictures and stacking them on top of each other to make up for the poor performance of the small hardware, denoising, and then applying a mask to subjects to lift shadows and adjust skin tones. This is the same kind of edit a photographer would do in Lightroom. It's not straight up AI generating content in the image.
Sadly Android phones are willing to pull cheap tricks to show better numbers or look better in some comparative youtube video so they pull this stuff.
Leica and Sony offer this as “Content Credentials” and anti-forgery digital signatures done on camera. It’s also been standard for many years for digital cameras used in criminal evidence procedures, Olympus tough series IIRC.
Me I just switched back to film. It’s about a buck a picture but I like the results I get, I’d rather have 30 photos than 300 to look through anyway.
Apple is apparently working on the latter, although it'll be opt-in and it's unclear whether it will also reduce the processing (I hope it does!): https://www.macrumors.com/2026/08/10/ios-27-apple-reference-...
This is a feature a lot of companies are looking in to right now. It's less about reducing processing and more about proving something is real and not AI generated. The iphone processing doesn't straight up generate things that aren't real.
It's also something I've been looking into. I've completely broken Google's implementation and I can sign any file as if it were real, more details will be public in the coming weeks.
Signatures are nice, but the contents of the image still matter. Any signs of manipulation should still be treated with suspicion, even if they're "legitimate" edits. The best way to avoid such signs is to have the bare minimum processing.
The technology is fundamentally flawed. It's essentially DRM that relies on making the signing key hard to access.
This is also my opinion. However, I still think it's worth raising the bar for plausible fakes. Minimising processing is another way to raise that bar (or at least, lower the floor). It's also not something you can do by default, because consumers demand image processing.
On Android, you can use OpenCamera. Optionally save in RAW and process in desktop RAW software.
Alternatively, buy a mirrorless or DSLR with some decent lenses.
How does that watermark feature work if you can just point the camera at an AI-generated image?
One of the proposals I saw was using the lidar to capture a depth map. You could also embed the GPS location.
If it's watermarked then it's by definition inauthentic. Watermarking is in-band signal manipulation.
Telling if a photo or a video is "true" is the same as telling if the written text is "true". You can not tell it by analyzing characters / words in a text, or pixels in a video.
You should analyze the subject who is providing that text / video and decide for yourself, whether you trust that subject or not.
E.g. "Elon Musk is 10 feet tall" - are you reading it on an anonymous X account or on a Business Insider social account? An attached photo of Elon Musk being 10 feet tall does not make a difference.
Eh, there are plenty of third party camera apps which do that (See: Halide and their Prozess Zero). I doubt a built-in one will go that way.
Well, they do their best, they don't do any userspace post-processing. They can't really help whatever the firmware in the camera and/or the camera driver are already doing.
It's not really possible to have zero processing of an image. At some point you have to interpret the sensor data and turn it in to a JPEG. The line people decide an image is "processed" is pretty subjective though. It's somewhere between the camera deciding what the white balance / exposure is, and AI replacing the sun with a moon.
The line I draw is at semantic editing. Manipulating images based on abstract geometric or statistical features like edges or color histograms is acceptable. Attempting to guess what those features mean is not. Traditional sharpening, color balance correction, focus stacking, lens distortion correction, etc. is fine. Red eye removal, skin tone correction, traditional dodging and burning, replacing a detected moon with a higher resolution photo, etc. is not. Unlabeled semantic editing is dishonest. AI image processing does not distinguish between geometric and semantic features, so it's always unacceptable, even for simple tasks like sharpening.
The one exception I make is cropping, because photographers have always had the ability to choose where to point the camera. Cropping based on meaning is not dishonest because it's inherent in the process of photography.
raw images exist
You can't see a raw image though. Any time you look at an image, decisions had to be made about how to interpret that raw image.
It's still not a jpeg
Any file format or data structure you can display on the screen is processed. The whole point is the debate around phone processing is not as clear cut as photos being processed vs unprocessed.
only because you're arguing in bad faith. The layperson does not consider the result of standard sensor debayering to be "processed".
I am wondering if the AI enhancement should be opt-in rather than the default and saving the original as well. Even if you want AI enhancement, future models are likely to be better.
Interestingly, one of their AI filters is actually called "Super Moon":
4.14.3 Camera Image Optimization
The camera image optimization AI algorithm is mainly used in the following camera features: Super Moon, AI Camera, Document Mode, Front Portrait Mode with Background Blurring, and Beautify.
https://trust.mi.com/docs/miui-privacy-white-paper-global/4/...
I still struggle with the idea that the chinese olympics had both CGI fireworks and real firework footage mixed together.
I cant fathom how this was a "if they know it doesnt matter at all" situation.
A complete acceptance of real / fake, who cares its about stuff looking great.
I mean I can be told it but I cant understand why you would waste millions on real works just to have fake ones, why not just go all fake? If its a half way house then does no one care what is real?
IE on linkedin somany AI videos with thousands of comments and likes and I become lost on ... do most poeple just not care if its real or not? I dont think they do at all. Am I mad to care?
Its very easy to say well everyone uses filters and they are not real. And its true! But somewhere and I have no idea where, there seems tobe a line for me where I say no, thats not cricket, you cant just replace part of an image with a template of something else. I have no idea tho any thoughts appreciated.
> I mean I can be told it but I cant understand why you would waste millions on real works just to have fake ones, why not just go all fake? If its a half way house then does no one care what is real?
I work with VFX and try to always have something real in the shot. Explosions and the like are easier to expand but quite difficult to just add to empty shot. Fireworks are a bit easier but there is interplay with the environment which is hard to fake.
All these films where they announce ”we did it for real with practical effects” usually mean that it’s a mix and match between practical and computer generated effects. It can look better than either on their own.
Also I guess they had audience there so you need something for them as well.
> why not just go all fake?
Perhaps they needed some real fireworks for those there in person, and to be able to claim that there really was a fireworks show, but wanted to amp it up for those viewing the event virtually (most of the world).
Does it come from a cultural mindset, or are people from China just as confused?
From what I have seen, the attitude towards AI in China is entirely different. The population largely believes it's a positive technology that will help improve their lives. While in America there is a seething hatred towards AI where people believe the benefits will all go to the billionaires while they are put out of a job.
If you look at the way Chinese AI companies act compared to American ones, it's easy to see how people feel this way.
The crucial difference: in China, if a billionare might become more powerful than the government, they get disappeared and re-educated: https://www.bbc.com/news/technology-56448688
Jack Ma is as scummy as the American tech elite, The whole 996 work schedule was from his companies. It's honestly good for society that people like him get reined in.
And it's not like they sent him to some north korean death camp. He's still one of the richest people in China, he just failed to deregulate the finance industry.
I wonder if it gets the moon the right way up in the southern hemisphere
i would not underestimate the tomfoolery Xiaomi is doing to give the impression their phone can take a zoomed in photo of the moon. It's pretty tasteless in my opinion
I got some amazing photos through a set of eclipse viewing glasses on my Vivo x300 pro. If I'd been a bit more organized and set up a stand and taped the glasses onto the phone I think I could have taken a video of the whole thing in incredible quality. But I just held the glasses in front of the lens and it worked.
What I found though was that I had to use the pro mode (which I use for basically everything anyway, to avoid exactly the kind of issues this article is talking about).
I did try on normal photo mode and I think what was happening was that as soon as I covered one lens with the glasses it switched to another.
It remains incredibly funny that they never turned off this "fake moon" feature after being caught the first time
I think you're confusing them with samsung who was caught in 2023 for this
https://www.theverge.com/2023/3/13/23637401/samsung-fake-moo...
You are right, but I think my point stands - anyone getting caught for this ought to make the others think twice
I don't think it's intentional. This behavior is a straightforward consequence of training something on pairs of "real" and "degraded" images. It learns what images look like and does the best it can. Philosophically, replacing the moon with a better moon is no different than any other computational photography technique that uses world knowledge, like reducing red eye from flash by detecting faces, or enhancing blurry text using knowledge of the alphabet.
The problem is philosophical - people are upset by certain types of enhancement error, but not others, and they can't even articulate the difference to themselves, let alone to an AI.
> people are upset by certain types of enhancement error, but not others, and they can't even articulate the difference to themselves
People are upset by enhancement errors that produce an image that is meaningfully different from the "true image". Replacing the sun with the moon or replacing the letters with different letters (e.g. https://www.dkriesel.com/en/blog/2013/0802_xerox-workcentres... ).
The philosophical problem here is that you can't get at the true image, you can only imagine ideas like "what the human originally taking the photo saw with their naked eye".
What constitutes a "meaningful" difference? If you replace flash-induced red eyes with a more plausible (yet inaccurate) eye color, people are perfectly happy with that, even though the camera indeed saw red photons. They're not happy if you run the exact same algorithm on a photo of someone with genuinely red eyes. It's not always about the "true photons" - the cause of the red eye is relevant, and so is the intent of taking the photo.
"what the human originally taking the photo saw with their naked eye" is a tempting standard, but a misleading one - there might not be a human eye near the camera at all, and at any rate there certainly isn't one co-located with the focal point. And shall we have cameras that render only a few degrees of high resolution color, and have a large-ish blind spot? It's what the human eye sees.
Swiftcoder can't possibly be expected to keep track of the difference between two Asian companies. They all look the same!
they will probably just add a condition into code to temporary disable it during eclipses
Because nobody cares. We live in a post-truth era.
the title should be "lies, damn lies, and photo enhancements"
Living in a post-factual world.