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.
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.
The reborn Nokia/HMD was also doing multi-sensor image stacking back in 2019 on the Nokia 9 Pureview.
Google has some great blogs about their camera technology https://research.google/blog/night-sight-seeing-in-the-dark-...
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.
From the authors of Halide.
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...
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.
https://www.google.com/search?client=firefox-b-d&q=ARGUS+pho...
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.
Subtle movement between shots can be used to create an image sharper than any of the inputs.
Phil introduced it with iPhone 11 as "deep fusion"
I’m guessing it has to do with how intensely focused / scared my brain was with the little guy
"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?
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...
So the ENHANCE is no longer just some lazy movie writing!
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.
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?
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.
A service will allow you upload a file and attach a date and signature. Additionally, you add the email addresses of anyone else that needs to also date and sign. The service then verifies all signatures and then dates and signs. Then it makes a SHA-512 hash, publishes it next to the signatures on its site and then publishes a url including the hash on the blockchain. When someone visits that link, they can see the hash of the file, dates, signatures and a final hash of everything. Anyone wishing to view the actual file must be able to prove their identity using their private key (or alternatively, host an encrypted version for each signer). Thus anyone can prove a few things:
1. You agreed that the document was correct on the date you signed. You can't date your signature any earlier than when the file was uploaded. This prevents you from uploading the wrong version of a document or predating your signature well in the past. It's like signing in front of a notary.
2. The additional person can verify the document is correct (no contract clauses have been edited, a photo is legit, etc). This gives credence to what you've posted, especially important for a contract.
3. The service verifies that the file you uploaded matches what everyone signed and the date everyone supplied is in the past. A requirement could be that everyone involved needs to have their public keys published on a well-known key server that at least verifies email addresses. A better option would be one that verifies through physical identity documents. Even better would be if your national ID card had PKI keys. Regardless, the service ensures that each signature has been tied to some sort of identity and that the signatures are all valid.
4. The hash published on the blockchain again can confirm that at the time of publishing, everyone agreed to the contents of the file and that everyone signed when they did. If someone claims the file was changed, we can simply verify the signatures and as long as no one was impersonated, we can prove that signers A, B, and C all signed a copy of the same document on dates X, Y, and Z.
5. The original document is still available for signer that still has their private key, even if it's expired. Additionally, the service has their own copy of the file that can be subpoenaed if necessary.
The entire system ensures a file has not been changed in any way since it was first published (via the signatures), real people have arrested to its truth (via public key verification), and that the attestation has not changed (via publishing the hash of all signatures on the blockchain). Additionally, nothing can be deleted without signed deletion requests from all signers.
You could potentially do this without the neutral 3rd party but then there's no verification when you actually attested to its truth, only that you published that attestation on the blockchain publication date.
Jory Miller?
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.
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.