Photos in iOS 27: the magic eraser stopped being a joke

Apple introduced three Apple Intelligence-powered editing tools for the Photos app in iOS 27 at WWDC 2026: a rebuilt version of Clean Up, the new Extend, and the new Spatial Reframing. The obvious headline is that object removal has finally caught up with Google and Samsung. But there are two more interesting points. Clean Up stopped embarrassing itself, yes, but the truly new feature is Spatial Reframing. And underneath all three, Apple made a decision that sets it apart from the rest: it intentionally put the brakes on its own tools.

Clean Up stopped embarrassing itself

Clean Up arrived with iOS 18.1 in late 2024, and it was never up to par. It worked for removing small things on simple backgrounds, but it broke down as soon as the scene got complicated. The typical case was trying to remove an object resting on someone’s head: in iOS 26, the tool distorted it instead of removing it. Google and Samsung were already doing this well.

In iOS 27, the leap is real. Apple rebuilt the models and promises more realistic fill even in complex scenes, and independent testing confirms it: it can now erase photobombers and reconstruct backgrounds with results close to Samsung’s, though still not quite matching it.

The less-discussed change is how it does it. Clean Up now offers three modes. Fast runs on-device and prioritizes speed for simple touch-ups. High Quality sends the image to Private Cloud Compute, Apple’s servers, for a more detailed reconstruction. Auto lets the system choose. The detail that matters: the best result comes from the cloud, not local processing. A good part of Clean Up’s magic depends on leaving the phone.

Spatial Reframing, the only truly new feature

Of the three, Spatial Reframing is the one that didn’t exist in any form. It changes the camera angle of a photo after it has been taken: you drag the image with your finger and the perspective shifts, as if you had shot it from somewhere else. Apple described it as going back in time to reframe the shot.

Internally, it works in three stages. First it scans the photo and builds a 3D model of the scene from the depth map, the same data the iPhone captures for Portrait mode. Then it lets you move a virtual camera through that model with a live preview. Only at the end does it generate new pixels, and only in the areas exposed by the perspective change; the rest of the image is left untouched. The foundation is the spatial modeling Apple developed for Vision Pro, which is why it also works with old photos and with shots taken on other phones, not just iPhone.

It has a clear usage limit: with small corrections, the result is flawless, but if you push the angle too far, faces and bodies deform. Used with restraint, it solves a concrete problem: instead of shooting the same scene five times from different angles to choose later, you take a single photo and adjust the framing afterward. Fewer bursts, less storage used.

Extend, fill with a limit

Extend does the opposite of cropping: it expands the photo beyond its original frame and fills in what’s missing with AI. It is useful for straightening a tilted horizon without losing the edges, changing the aspect ratio, or giving breathing room to a subject that ended up too close to the margin.

But Extend is not a free-for-all. Apple capped it at 25% per side and allows only one application per photo. You can’t stretch an image indefinitely until you invent half a scene. That cap is not accidental.

The limit is Apple’s stance

The three tools share one rule: AI only generates where it is needed. Spatial Reframing fills only the gaps opened by the perspective change, Extend has its 25% cap, and Clean Up reconstructs what you removed; it does not invent from scratch. That is the difference between refining a photo and fabricating it. Compared with the anything-goes generative approach from Google and Samsung, Apple chose to put up guardrails.

That stance has a weak point worth calling out. Apple executives keep using the word "refine," but changing the camera angle or the direction a subject is looking is generating new content, not refining it. And there is a technical nuance that clashes with the messaging: when a photo does not include a native depth map, Spatial Reframing calculates one algorithmically. That estimated depth is an approximation, not a record of what the camera saw. Apple’s argument that its edits are anchored in what the lens captured is less solid than it appears.

Which iPhone you need

The AI editing trio —the new Clean Up, Extend, and Spatial Reframing— requires Apple Intelligence, so it runs on iPhone 15 Pro or later. Some features launch only in English and in beta. The rest of Photos does not have that requirement: performance and organization improvements come to all iPhones compatible with iOS 27, starting with iPhone 11. The app opens camera roll photos up to 70% faster, a dedicated section for what you captured with the camera returns, photos can be tagged with keywords to make them easier to search, and temporary shared albums appear, deleting themselves after 30 days.

Conclusion

For anyone who just wants to erase an intruder and move on, Clean Up finally works: leave the mode on Auto and it will use the cloud when it makes sense. For anyone who agonizes over every frame and regrets how a shot came out, Spatial Reframing is the most useful part of the package, as long as you don’t force the angle until faces start to melt.

The bigger picture is more interesting than any of the three features. This is only part of what’s new in iOS 27, but it sets the direction: Apple is betting that AI editing with limits —refine, don’t fabricate— is its differentiator against Google and Samsung, which generate with almost no restraint. Whether that restraint reads as a principle or as falling behind is the question iOS 27 still has not answered.