upvote
Hijacking this comment because of this line that reminded me of something:

"There is a crazy amount of image processing going on behind the scenes in each smartphone"

I bought the Pixel 10 because it said it had great zooming capabilities. After 5X (that means 10X), it's completely stupid. Basically it takes a very blurred photo and then tries to reconstruct it with AI. The final results are completely awful and really distant from reality.

Any advice in how to improve that?

reply
You can use third party camera apps - https://f-droid.org/en/packages/net.sourceforge.opencamera/ is a decent one. I know there's a few professional ish options too with dslr ish focus control.
reply
In photos after you can toggle between the AI modified and original. The original isn't as blurry as the preview shown - there's still a lot of image stacking happening and it's often very decent in my experience.

Edited to add: I have a pixel 10 pro which has a better zoom lens, so I could be having a different experience than you...

reply
One solution is to buy a more expensive Pro version that has optical 5x zoom (or, rather, a third camera that start at 5x).
reply
Also have a 10 pro .. but the AI-enhanced zoom goes up to 100x and is actually pretty handy for working out what tiny things in the distance actually are ... even if it's only a representation of that thing. Often use it for identifying birds, boats, etc.
reply
I think my p10p's 100x zoom is neat, but I'd strongly caution against using it for ID in any sense, or necessarily even treating it as a photograph. It will "straight up hallucinate" objects into existence, more of a creative interpretation of 100x digital zoom than an oracle into the identity of something far off.

Makes for a fine image if "looking good" is the objective (which it often is), but I wouldn't trust it to disambiguate anything uncertain.

reply
> Basically it takes a very blurred photo and then tries to reconstruct it with AI

So the ENHANCE is no longer just some lazy movie writing!

reply
Even funnier:

The biggest of the "sharpening with AI" models, if you let 'em go, they'll turn you into a celebrity on accident.

If you go to the Topaz AI forums on Facebook, you'll see examples of where the AI sharpening tool went off the rails and turned a blurry photo of someone (typically the user who bought the software, who is trying to clean up their own photos) into whichever celebrity that is in the AI model who resembles them.

IE, it might turn a photo of YOU into a photo of Robert Downey Jr.

reply
A couple years ago, we needed a good photo of my grandfather for a photo collage. The best photo I had of him was where he's sitting next to a very decrepit looking old woman (some family friend?). It really was a great composition, great lighting, background and pose but the old lady was killing the vibe. My wife is really skilled at Photoshop but removing this old lady would have been really difficult so we decided to try the generative fill, something she had never used before. She selected the lady and typed in "delete this person". Instead of deleting her, it creates this strange, human like being. We try again and again with different permutations of the same command. Most of the attempts just replace the old lady with a small asian child. Eventually we just try with a blank prompt and it gets it in the first go. The Asian child thing was really strange, it truly looked like the same kid but just slightly different appearance each time.
reply
> it truly looked like the same kid but just slightly different appearance each time

Jory Miller?

reply
That's hilarious!

Also wonder how how long until photos are no longer court-admissible evidence because they get overwhelmed with the burden of the amount (and the cost) of the "non-AI" certification?

reply
I can see film and instant cameras making a comeback for anything that needs to be used as evidence.
reply
We've been living in the age of photoshopped photos for decades now. You can't just present a random photo as evidence without being able to back up its provenance.
reply
As long as you know how to edit metadata, you can definitely fool minor “trust systems” like adjusters, police, etc. It’s a big gap. (For any future readers, no I have not done this and would never do it, really.)

The only solution I see is some kind of hardware-linked signature showing that the image was produced at a certain time, by a certain device. In theory, you could make this signature somewhat “provable” and anchored to a specific time by combining something like 1) the latest Bitcoin block hash, and 2) a blockchain anchor, thereby proving that the photo was taken within a certain time widow after the hash was known and before the anchor was published. But even that arrangement is highly susceptible to hardware attacks.

reply
I'd like to be able to use my hardware security key for this purpose (the hardware-linked signature)
reply
The ENHANCE meme used to drive me nuts because it is clearly impossible. My brother and I even briefly competed to find the earliest ENHANCE reference (I no longer remember what we concluded, but it goes surprisingly far back). The fact that it actually exists now is mind-boggling (even though it is, by necessity, still a fiction).
reply
Can't wait for the first use of such images as evidence in court.
reply
Wasn't that the same manufacturer that got caught with the moon AI stuff? Lol
reply
reply
Gotcha, thanks! Funny that they didn't learn from that one
reply
That was samsung
reply
No, 10X is fine - there is enough resolution on the sensor to just crop the 5x images to 10x without any weirdness.

After 10x it's all over. Which is in itself quite insane. The camera is truly amazing, and complaining that this 10mm thick camera is no good at further than 10x seems utterly spoiled and ridiculous to me.

reply
go closer irl
reply
Yes exactly, the site shows photos all taken from within the app. I have done A/B tests against photos taken in Camera.app and the Photosynthesis composite does significantly beat the detail central area covered by the second lens in Photosynthesis.

Because of the limitations of capturing multiple lenses at once through isVirtualDeviceConstituentPhotoDeliveryEnabled, there is a second capture mode in the app for 'Quality' which takes a simultaneous capture then takes sequential shots from each lens which do benefit from Deep Fusion, then uses those for the stitch (as long as the Deep Fusion images outperform the simultaneous capture results). That's also what I do for our Night mode.

reply
Phones have been doing this for a LONG time, if you had a oneplus 7 pro (released may 2019) and took it into an extremely dark area, like where your eyes can't see anything at all, and take a 'night mode' photo, it asks you to hold it steady for about 4-5 seconds and is clearly taking a series of photos with long exposure time. Then it stacks them and uses data from the sensor of hand movement jitter to deblur.

The result when done outdoors (like, standing in a dark forest at midnight) will be a photo where you can actually see what's around you, and details show up as if it was a brightly full moon lit night.

reply
That's a bunch of photos from the same camera, not 2 lenses at the same time
reply
I know, I just don't think it's a huge jump from "take multiple photos over a 2s period from one camera and stack them" to "take several photos from discrete physical cameras and stack them". As a subscription of paid app I am not the target market for the person who made this, I agree with another person commenting in here that said they should be working for a phone manufacturer.
reply
I’d argue the opposite. The geometry involved in reconciling photos from two separate lenses with different focal lengths is quite challenging. If you pay close attention you will see some differences in the before/after samples in the site (for example the ear in that kitchen selfie).

The math is surely more complex than aligning same-lens photography even when you consider the change in angle and perspective between two shots from the same lens.

reply
Not really. Whether you have two different lenses or the same lenses with some offset from hand movement you need to accurately model the lens(es). Once you can do that for one lens you can just as easily do it for two - it's just more variables to optimize, not a fundamentally harder problem. Panorama software has supported images from different lenses since forever.
reply
The Amazon Fire Phone did this in 2014. I'm not aware of any prior art in mainstream devices aside from gimmicks like 3D (stereoscopic) photos. I can't remember if the Fire Phone had any relevant gimmicks but it certainly used multiple lenses to produce a single image.
reply
UAV sensor people made a thing that uses 300+ phone size sensors in a single pod, and software workflow to fuse the output together. It's been around quite a few years now.

https://www.google.com/search?client=firefox-b-d&q=ARGUS+pho...

reply
Ah yes, how could I forget the ARGUS-IS, a household staple of the early 21st century. Those were the days...
reply
I keep an extra one next to my panini press, the image quality from the whiteboard notes area on my fridge has never been sharper!
reply
I think it's interesting that the same functionality appeared on iPhone in 2019 as well. Google tells me it was the iPhone 11 and iOS 13.
reply
Night Sight appeared on the Google Pixel camera app in 2018. I think Nokia was doing it first on phones before Microsoft destroyed the remains of their phone business.

The reborn Nokia/HMD was also doing multi-sensor image stacking back in 2019 on the Nokia 9 Pureview.

reply
The related paper (2019): https://google.github.io/night-sight/

Google has some great blogs about their camera technology https://research.google/blog/night-sight-seeing-in-the-dark-...

reply
There is an iOS app that does this called Spectre: https://spectre.cam

From the authors of Halide.

reply
It's different from just stacking exposures, it's a paid app that does progressive animation of exposures for creative effect and says it's "AI-powered" whatever that means.
reply
It's a paid app that I literally have, and all it does is take multiple photos and composite them together. It can either naïvely average them together in-place like a traditional long exposure, or align them using machine learning. I'm actually appalled they use the term "AI" considering who they are but that doesn't matter, it's ML. I don't use any of the animation crap, but the aligned long exposures can be useful for ultrawide takes since that camera is so noisy.

Edit: actually I can't seem to find the setting that aligns them. I seem to remember it being able to do that but maybe it doesn't... which would be disappointing...

reply
I have it too but after like 3 times I never used it again, felt like a cheap gimmick. I regret falling for hype. If anything they do good marketing
reply
> I'm actually appalled they use the term "AI" considering who they are but that doesn't matter, it's ML.

I agree, and I built it!

The app launched in 2019. Ideally I would have said ML, but it felt like AI had more recognition among non-technical users.

If I launched it in 2026, it would 100% say "ML" not only because it's more technically accurate, but to make it abundantly clear it doesn't generate slop.

reply
People are slapping "AI" on anything these days to try to sell more units of whatever it is.
reply
I use iPhones (paid for by work) as a tool for my agronomy job. I take a large amount of photos throughout the day to help document what I find in the field.

Since at least the iPhone 12, holding the phone still and letting scene sit for a few seconds will result in a sharper photo. I’m not sure if it’s a subtle version focus stacking or a sharpening algorithm, but it’s definitely there.

On my 16e, immediately looking at a photo just taken can result in the image noticeably sharpening after a second or 2, so this may be a post processing step as well.

reply
It's probably a version of this: https://dl.acm.org/doi/abs/10.1145/3306346.3323024

Subtle movement between shots can be used to create an image sharper than any of the inputs.

reply
From what I can tell, it's taking multiple photos and stacking them in post. So giving the phone more time to capture sharp exposures to stack would do a lot to help.
reply
Reading title and this ^ comment before loading the page made my brain assume that this is about combining front and rear image sensors, and additionally using eye/facial tracking to infer user intention + satisfaction to help with stabilization/framing/post-processing. I'm glad(?) we aren't there yet.
reply
I think this is just Deep Fusion.
reply
deleted
reply
One time I took a photo of a snake in the road with my iphone in low light. When I looked at the photo afterwards the snake was removed by the AI enhancement
reply
Similar story but my low-light snake photo turned out much less “scary” than in real life.

I’m guessing it has to do with how intensely focused / scared my brain was with the little guy

reply
But they do do this? They talk about it? It's not a secret
reply
It is part of Apple’s “secret sauce”, yes.
reply
Thank you for confirming my suspicions, lol
reply
Don't they do this with the "macro" setting already ?
reply
This should be easy to check. In these photos, the center of the image has higher detail than the edges.
reply
Maybe on this guy’s app, but not the way Apple does it.
reply
If that's the case then it would disprove the original assertion that Apple already does the same thing. You can't have the same high detail at the edges without making shit up because those parts of the image are out of frame for the higher zoom lens.
reply