Google Photos’ Auto Frame lets you reframe your photos after the fact

We’ve all been there: you snap a photo, look at it later, and think “if only I’d moved a few inches to the left” or “the angle is just off.” Cropping doesn’t fix it—you’re still stuck with the same perspective, just tighter. Google’s new Auto Frame feature in Google Photos aims to solve that by letting you re-angle the shot after the fact.

It’s now live in the app, and it’s not just a fancy crop. The system actually interprets your 2D photo as a 3D scene, figures out where the camera was, and then generates what would be visible if you’d shot from a different position. Think of it as a time machine for your camera roll.

How it works

The approach, detailed in a blog post by Google’s Marcos Seefelder and Pedro Velez, breaks down into two stages. First, a 3D point map estimation model reconstructs the scene geometry, paying special attention to human bodies and faces to avoid distorting identities. It also guesses the original focal length.

Then, classical 3D rendering shifts the virtual camera to your new desired angle. But that alone leaves holes—areas that were hidden behind the subject. That’s where generative AI comes in. A latent diffusion model, trained on pairs of images with known camera parameters, fills in those gaps. It’s not just guessing; it’s inferring what should be there based on the scene layout.

The result is a re-composed image that looks like you actually took a different shot. In my testing, it handles simple scenes well—portraits with plain backgrounds, landscapes with clear depth. Busy scenes with lots of overlapping objects can get messy, but that’s expected for a first-gen feature.

Why this matters

Most generative image editing tools today are about changing content—swap a background, add an object, remove a person. This is different. It’s about changing the geometry of the capture itself. That’s a genuinely hard problem because you’re not just painting over pixels; you’re reconstructing a 3D space from a single 2D image.

Google’s trick is decoupling the 3D estimation from the image generation. That gives them control over camera intrinsics (focal length) and extrinsics (position, orientation) separately. Classic photogrammetry approaches would struggle with the incomplete data, but the diffusion model fills the gaps convincingly.

Practical uses

The obvious use case is fixing that selfie where the wide-angle lens made your face look weird. Auto Frame can adjust the perspective to something more natural. Group shots where someone’s face is partially hidden? Move the virtual camera slightly and the model fills in the missing cheekbone.

It also works for landscapes—shifting the horizon or revealing more of a foreground element. But don’t expect miracles. If the original photo doesn’t have enough context (like a close-up of a texture), the model has nothing to work with and the result looks like a blurry guess.

The catch

It’s only available in Google Photos on Pixel devices for now, and it’s part of the Auto Frame feature, not a standalone tool. That means you can’t manually set the angle—the system suggests a recomposition automatically. You can accept or reject it. I’d prefer manual control, but I get why Google kept it simple for launch.

Also, it’s not real-time. Processing takes a few seconds on-device, which is fine for a single photo but annoying if you’re batch-editing. And since it’s ML-based, results vary. Faces are generally well-preserved, but I’ve seen it add weird artifacts to hair and eyeglass reflections.

Bottom line

Auto Frame is a genuinely novel approach to photo editing. It’s not just another filter or generative fill—it’s rethinking what a photo can be after the shutter closes. I expect we’ll see this kind of spatial recomposition become standard in camera apps within a few years. For now, it’s a neat trick that works more often than it fails, and that’s impressive for a first release.