Google Photos Can Now Re-Angle Your Photos After You Take Them

I’ve been playing with the new Auto Frame feature in Google Photos, and it’s one of those rare tools that actually does what it promises without feeling like a parlor trick.

We’ve all been there: you snap a selfie, the smile is perfect, but the wide-angle lens makes your face look distorted. Or you catch a great moment, but the angle is slightly off—maybe you wish you’d shot from a lower position or captured more of one side of a subject. Classic editing tools let you crop or zoom, but that doesn’t change the underlying perspective. Zooming in doesn’t alter parallax, and cropping can’t show you what was just outside the frame.

The team at Google (Marcos Seefelder and Pedro Velez from DeepMind) built this thing that treats a 2D photo as a frozen 3D moment. Instead of just cutting edges, it actually re-positions a virtual camera within that 3D space and generates the missing content. The result is an image that looks like it was taken from a different angle entirely.

How it actually works

Most generative image editing tools try to do everything in one go—prompt goes in, new image comes out. This approach is different, and honestly, smarter. It splits the problem into two distinct stages.

First, a 3D point map estimation model analyzes the original photo. For every pixel, it figures out where that surface patch sits in 3D space. The model is specifically tuned to handle human bodies and faces well, which matters because bad reconstruction there would break identity preservation. It also estimates the original camera’s focal length.

Second, it uses classical 3D rendering to project that point map as if seen from a new camera position and orientation. You can change both the camera pose and the focal length—full control over the virtual lens.

But here’s the catch: moving a virtual camera reveals gaps. If you orbit around a subject, you expose parts of the background that were never captured. The point map is incomplete. So the system uses a generative latent diffusion model to fill in those holes. This model was trained on image pairs with known camera parameters—it learned to take a re-rendered view and reconstruct what the second image should look like.

The diffusion model also handles retouching. It smooths out artifacts from the 3D projection and ensures the lighting and textures feel natural. At inference time, it uses classifier guidance with regional scaling to balance fidelity to the original with plausible generation.

What it feels like in practice

I tested it on a few photos from my camera roll. The results vary depending on how much the angle changes. Small adjustments—tilting up slightly, shifting left a bit—look nearly seamless. Larger moves, like trying to see around a person’s head, show more generative artifacts. The model is conservative, which I appreciate. It doesn’t invent wild details; it fills gaps with plausible textures.

The feature is live now in Google Photos under Auto frame. It’s automatic—you don’t dial in camera coordinates. The system suggests a new perspective based on scene understanding. You can accept it or tweak further.

Is it perfect? No. Complex scenes with lots of overlapping objects can confuse the 3D estimation. And the generative inpainting sometimes produces blurry patches in areas where the model isn’t confident. But for a first public release, this is impressive.

Why this matters

What’s interesting here is the decoupling of 3D estimation from image generation. Most competitors try to do end-to-end diffusion, which gives you less control and more hallucinations. By separating the geometry step, Google gets a more predictable output. The 3D point map acts as a structural constraint, so the generative model only fills in what’s missing rather than reimagining the whole scene.

This also means the approach can be extended. Once you have a 3D point map, you could theoretically change lighting, insert objects with correct perspective, or even animate the camera path. The foundation is solid.

I’ve seen a lot of “AI photo editing” features that are just fancy filters or upscalers. This one actually changes the geometry of the captured moment. That’s a genuinely new capability.

The catch

It only works on single images, not video. And it requires a modern device with enough compute to run the models locally—Google isn’t sending your photos to a server for this. The models run on-device, which is good for privacy but means older phones might not get the feature.

Also, don’t expect cinematic camera moves. The angle changes are subtle, meant to fix composition rather than create dramatic new perspectives. If you want to fly around a scene, you need a multi-view capture system, not a single photo.

Still, for fixing that one photo where everything was right except the angle, this is a genuinely useful tool. I’ve already salvaged a few shots I was ready to delete.