Google Photos Auto Frame rewrites perspective in 2D shots
A two-stage ML pipeline now lets Google Photos change the camera angle of a photo after it was taken, not just crop or zoom.
Can Google Photos change the camera angle of a photo after you've taken it?
Yes. Google has shipped a feature called Auto Frame inside Google Photos that reconstructs a photo as a 3D scene and then re-renders it from a new camera position. The result goes beyond cropping or zooming: parallax relationships between objects actually shift, and content that was never in the original frame can be generated.
Why cropping wasn't enough
Classic editing tools operate on a flat image. Zooming in doesn't alter the spatial relationships between foreground and background objects, and cropping can't reveal anything outside the original field of view. The team's stated motivation was the common experience of photos that are almost right — a face caught at a slightly unflattering angle from a wide-angle lens, or a composition where the subject is slightly off-center.
How the two-stage pipeline works
The approach separates the problem into two distinct steps: 3D scene and camera estimation, followed by generative inpainting and retouching. Keeping these stages decoupled lets the system manipulate camera intrinsics (focal length, field of view) and extrinsics (position and orientation in space) independently before any pixels are generated.
In the first stage, an internal 3D point map model analyzes every pixel of the original image and estimates the 3D surface it represents. The team notes the model was specifically configured to reconstruct human bodies and faces accurately, to reduce artifacts that could distort a person's appearance or identity.
In the second stage, generative AI fills in regions of the scene that were never captured by the original camera. This is what allows the feature to show content just outside the original frame rather than simply stretching or mirroring edge pixels.
Automatic camera suggestion
Users don't have to manually dial in new camera coordinates. ML models analyze the scene contents and propose new camera parameters automatically. The system infers the original camera position from the image's spatial layout and then decides how to reposition it — which is how the feature surfaces as an automated suggestion inside the Photos app rather than a manual 3D editing interface.
Identity preservation as a design constraint
The explicit tuning of the point map model for faces and bodies signals that the engineers treated portrait distortion as a primary failure mode. Wide-angle selfie distortion — the kind that makes facial features appear stretched or misaligned — was called out directly as a use case the feature is meant to address. By faithfully reconstructing geometry before handing off to the generative stage, the pipeline aims to keep a subject recognizable across viewpoints.
What this is not
Auto Frame is not a general-purpose 3D scene editor. Users are not presented with free camera controls or a depth map to manipulate. The feature surfaces recommendations automatically and is integrated into the existing Google Photos editing flow. It also doesn't work from video or multi-image inputs based on information in the announcement; it operates on a single still image.
Availability
The feature is described as live inside Google Photos as part of the Auto Frame feature. The announcement was published on April 22, 2026, and was authored by Marcos Seefelder, Staff Software Engineer at Platforms & Devices, and Pedro Velez, Senior Research Engineer at Google DeepMind.
The core technical distinction worth noting: most generative photo editing tools add, remove, or restyle content inside the original viewpoint. This pipeline changes the viewpoint itself, using generation only to fill the gaps that the new angle exposes. That's a different problem than inpainting a removed object, and the two-stage architecture reflects it.