iOS 27 is Apple’s most AI-centric release to date, introducing the highly anticipated Siri chatbot and other smart features across the system. One of the notable additions is the AI editor in the built-in Photos app, which features the Reframe, Extend, and Clean Up tools.
To test how effective these AI-powered image manipulators are, I applied them to a wide range of photos I’ve taken with my iPhone 17 Pro Max. Here are some of the great and not-so-great results.
In iOS 26, Apple introduced Spatial Scenes, a Photos app feature that detects subjects and adds a 3D effect that reacts to your iPhone’s movements. iOS 27’s new Reframe tool similarly analyzes the scene and allows you to tweak the angle or composition of your shot. Once set, the AI model will regenerate your image with the updated point of view.
To minimize hallucinations, Reframe limits perspective changes to areas it can comprehend, so you may be locked to a specific direction when reframing a certain image. Additionally, by default, you cannot repeatedly reframe an image in the same direction, but I was able to bypass this limitation by saving reframed photos to Files and then saving them again to the Photos app.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third).
For the first test, I reframed a photo of Beirut’s corniche twice. I was pleasantly surprised by how accurately it changed the handrail’s perspective. It also maintained the rainbow’s arc and correctly generated the rest of the rocky shore. It did, however, incorrectly move the trash cans and streetlights away from the handrail.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third).
Next, I reframed a photo of an iced latte overlooking the city of Byblos. The first edit was highly realistic, preserving the texture and shape of the ice cubes. The second edit similarly generated parts of the wooden table in a very realistic manner, although the distant background looked clearly artificial.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third). The third test was exceptionally disappointing, as the AI model seemingly failed to recognize a regular airplane. The first edit bent the aircraft, while the second incorporated irrelevant components.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third).
For the fourth test, I reframed a Ferris wheel at sunset, which the AI similarly failed to properly regenerate. The first edit wrecked its circular form and skipped generating new cabins where necessary. The second Reframe surprisingly fixed the circular shape (for the most part) but still didn’t add in the missing gondolas.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third).
Next, I headed to the mountains and reframed a lavender shrub looking over a village. The first edit was highly realistic and somehow enhanced the photo’s colors and added more clarity to the background. The second edit, on the other hand, completely ruined the shot through obvious AI hallucinations and pattern repetition.
Original, unedited photo (first); first attempt to use Reframe (second); another attempt at Reframe (third).
For the final test, I reframed myself blowing a fluffy dandelion in the middle of nowhere. The first edit failed to maintain the stem’s structure and made me look older. The second edit distorted my face and the dandelion’s stem further, but it did accurately generate the lake’s surface.
While several first edits look believable, I don’t see myself using the Reframe feature, as it doesn’t deliver the actual truth. Still, many users will appreciate the ability to effortlessly change the perspective of their shots.
Extend is essentially the opposite of cropping, letting you generate nonexistent parts of an image along its edges and expand it accordingly. Like Reframe, Extend restricts multiple image extensions to limit AI hallucinations, but I was similarly able to bypass it.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third).
For the first test, I headed to the Port of Tripoli, where I attempted to generate a cropped lifebuoy and the background boats towards the right. Unfortunately, Extend failed to recognize the object and limited my extensions to the image’s left edge, producing underwhelming results with both edits.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third).
For the next test, I had very low expectations, as it was a dimly lit photo of my car’s interior. To my surprise, the first extension correctly identified the infotainment system’s display and generated a relevant result. The second edit similarly estimated the car’s borders, accurately generating the left mirror and window.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third).
The third test was a tricky, close-up shot of cement boulders. Both extensions correctly detected the objects’ patterns and generated realistic results.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third). Next, I had Extend generate the bottom section of a cropped lighthouse. The AI precisely replicated the colors and pattern, including those of the clouds. The second extension correctly generated a lifelike sea, which I found impressive, but it also hallucinated a second extra-large lighthouse.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third).
For the fifth test, I attempted to extend the buildings on Jounieh’s hillside. In both edits, the AI model concealed the view by generating dull buildings on the left and right sides, and it inserted generic cars on its made-up road.
Original, unedited photo (first); first attempt to use Extend (second); another attempt at Extend (third).
Lastly, I extended a dark alley twice, and the feature maintained the original photo’s vibe by generating suitable cars and buildings.
Like most AI features, Extend’s results can be unpredictable. Sometimes It’ll fail to extend a well-lit shot yet somehow manage to get a complex setting right.
Clean Up in Apple Photos first launched with iOS 18, letting you remove unwanted people or objects in images. With iOS 27, Apple is upgrading Clean Up by letting you pick between Fast and High Quality edits. The former mode is comparable to Clean Up on iOS 18 and 26, letting you quickly remove clearly identifiable items. The latter mode analyzes selected items’ surroundings and generates more realistic edits, even in tricky shots. To evaluate Clean Up on iOS 27, I edited each photo using both modes separately.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third).
For the first test, I removed people visible on the cobblestone street. The Fast edit clearly looked like a manipulated image, while the High Quality one looked far more realistic.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third).
In the second test, I removed the cranes and people at the Port of Beirut. The Fast result similarly looked clearly AI-generated, while the High Quality one looked much cleaner. The difference is particularly obvious when inspecting the gavel’s edges and the Act for Justice text.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third). Next, I decided to Clean Up the buildings blocking my sunset view, minus the old lighthouse. The Fast mode expectedly failed miserably, while the High Quality one generated a usable result. While you could tell the latter edit is AI-generated, its precision is still admirable, given that it had to make up a significant portion of the shot.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third). The fourth test pushed Clean Up to its limits by attempting to remove a wide range of complex objects in front of two kittens. The first edit unsurprisingly did not pass, while the second actually managed to perfect the scene. It correctly generated the kittens’ bodies, along with convincing paws.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third). Clean Up 1.0 infamously failed to generate human faces. To see if the latest iteration of the feature addresses that, I hid my face with a camera and had Clean Up on iOS 27 handle it. The Fast mode behaved as expected, disfiguring my face with a digital whirl. Meanwhile, the High Quality mode successfully generated the covered parts of my face, but the end result looks nothing like me. While a welcome improvement, the AI model should ideally take on-device selfies into account. This would allow it to generate more realistic images of iPhone users.
Original, unedited photo (first); Fast Clean Up (second); High Quality Clean Up (third). Finally, I used Clean Up to clear my view as I tanned on the sandy beach. The Fast mode struggled to remove a couple of items, and its edits looked too artificial. Conversely, the High Quality mode handled the task professionally.
While the High Quality Clean Up mode consumes significantly more time and power, its results are infinitely more reliable. The Fast mode is fine for basic quick edits, like removing specks and blemishes.