Apple’s new AI photo editing tools mostly work, for better and worse

The most popular camera in the world just got its first set of serious AI photo editing features, and I don’t think any of us are ready.

As far as AI photo editing goes, the new features in iOS 27 are pretty tame compared to what you can do on, say, Google’s Pixel phones. But for the iPhone, they represent a tipping point in what the native photos app allows you to do to your photos. I mean memories. I mean, I don’t know anymore.

These new features are part of the iOS 27 developer beta right now, so bear in mind that Apple may continue making tweaks to them before they’re released to the general public. There are three, or maybe two and a half, new AI editing features in this update. The new Clean Up tool counts as half, because it existed before but was so bad it didn’t really count. That’s the tool that lets you take photobombers out of the background of your photos, and it got a major upgrade this year. There’s also Extend, which lets you expand the edges of your photo using AI to paint in some plausible-looking filler. And there’s Spatial Reframing, which mimics the effect of moving the camera around the scene to let you recompose an existing photo. It’s the most ambitious and maybe the most problematic of the three.

But first things first: Clean Up. It’s actually good now. Instead of only using on-device models to remove objects and fill in details, it can now use more powerful models in the cloud. This is what Google has been doing for years now, and it’s why the company’s Magic Editor tools were miles better than the version Apple introduced last year. That totally on-device Clean Up wasn’t very good at painting in convincing details to replace what it removed. It left weird artifacts and was generally more trouble than it was worth. Clean Up 2.0? It does the job.

Using AI to remove stuff from photos is the generative editing tool that I’m the least queasy about using. I’ll use it to remove a booger from my kid’s nose or take a stranger out of the background. This new version in iOS does all of that without a problem, and I think it’s going to be popular with iPhone owners.

Stepping up to the next level of complexity and what-is-a-photo-ness, there’s Extend. Think of it as cropping, but in reverse. It lets you expand the edges of your frame, which you might want to do if your composition was too tight on your subject and you want to give it a little more breathing room, for example. Extend lets you do this, but only to a point. It seems to avoid making edits to people, and will sometimes tell you a photo can only be extended in a particular direction. It will only add a little bit of padding, too, which minimizes the kind of shenanigans it can be used for. I appreciate that. Like Clean Up, it does its job convincingly. It seems predisposed to looking for symmetry, which usually works. It added part of a rally car that was out of frame in my original image, adding a side mirror to match the one already in the photo.

It doesn’t seem as eager to make stuff up to put into your photos as, say, Samsung’s early efforts. But I did catch it adding a potted plant on a side table; it looks reasonably convincing, but I know that it’s not a real plant. I’d feel weird about that if I were to put that photo on Instagram.

Extend works with your photo in a two-dimensional space; Spatial Reframing adds a third dimension. It builds on an existing feature that makes your photos look 3D-ish, allowing you to reframe a photo as if you had physically moved the camera and changed your perspective of the scene. You can’t go too far with it — only about as far as you could have moved your arm when originally taking the photo. But the idea is that you can fix your framing if you didn’t quite nail it when you took your shot.

Look, this appeals to my Type A nature. Sometimes I’ll love everything about a photo I took, except wouldn’t it be better if I had just stepped to the left to avoid framing something distracting in front of my subject? You can’t always catch that stuff in real time. These are the minor adjustments that Spatial Reframing is designed for.

Seems reasonable, except that there’s room to introduce existential chaos, even with minor adjustments. I tried changing the framing of a photo I took at a tech talk with Apple executives following the WWDC keynote. Since I was sitting off to the side, one of the execs onstage was mostly obscured in the original shot. I changed the framing and the AI kind of made up a guy sitting next to Craig Federighi.

Predicting what should come next in a two-dimensional space seems like an easier problem than in three dimensions. The results from Spatial Reframing compared to Extend reflect that. They’re weirder. The farther you are from a subject, the less latitude you have to “re-compose,” and the more realistic the AI-generated stuff is. But you end up with an image that’s only subtly different from the actual photo you took, and at that point, what are we doing?

In photos with a subject that’s closer to the camera, things get strange. The effect of “re-composing” is more dramatic, and the AI has to work harder to fill in the gaps. You can change the perspective in a selfie, but that means the AI has more detail it needs to fill in on your face, which gets a little uncanny valley pretty quick. Even within its limited adjustment range, it can make faces look a bit skewed and “off.” It’s more prone to inventing things that weren’t there. Sure, it sounds nice being able to rescue a photo from a composition that wasn’t quite right. In practice, I don’t think I like it.

It’s a small comfort that images edited with these AI tools get Synth ID labels signaling that they have been modified with AI. Instagram picked up on this info when I uploaded a couple of images, but only surfaces that information if you tap on the “AI Info” menu for that image. Labels aren’t exactly an airtight solution right now, and I think the bigger danger is the quickly eroding notion that we can usually trust a photo that someone takes and posts from their phone. Apple is certainly not the most audacious player here. But even introducing a little bit of doubt about the provenance of a houseplant on a side table, or whether someone was indeed standing exactly where they appear to have taken a photo from, can add up to a lot of trouble over time.

Photography by Allison Johnson / The Verge