It's all about the angle: Your photos, re-composed

Google just quietly solved one of photography's oldest annoyances: the shot that's almost right. You know the one — the selfie where the wide lens makes your nose look enormous, or the portrait where you wish you'd tilted the camera two inches lower. Cropping never fixed that, because cropping just trims a fixed viewpoint. It can't invent the part of the room that was outside the frame, and it definitely can't undo distortion baked in by the lens.

Google's new approach, folded into the Auto frame feature in Google Photos, treats a photo less like a flat image and more like a frozen 3D moment. The system builds a point-map estimate of the scene from a single picture, figuring out depth, surface shape and even the original focal length, with extra care taken to preserve faces and bodies accurately. From there it can literally move a virtual camera through that reconstructed space, adjusting both position and lens properties as if you'd been standing somewhere else when you pressed the shutter.

Moving the camera, though, opens gaps. Shift the viewpoint and you expose slices of background the original lens never captured — the classic problem with any 3D reprojection. Google's fix is a generative diffusion model trained specifically on pairs of images shot with known camera parameters, so it learns to plausibly paint in those newly revealed edges rather than just guessing. During inference, the model is guided region by region, staying faithful to what was actually photographed while improvising convincingly on what wasn't.

The more interesting part is that this whole pipeline runs without you touching a slider. Google trained separate models to detect face position and orientation, then uses that alongside the 3D map to pick what it considers the ideal framing automatically — pulling the virtual camera back to counteract the bulging, too-close-to-the-lens look that wide-angle front cameras are notorious for. The result shows up quietly as a second option inside Auto frame's suggested edits, a one-tap fix rather than a manual tool.

It's a real departure from typical AI photo editing, which mostly nudges pixels around in 2D. This treats the picture as evidence of a real physical scene and reasons about it geometrically before generating anything, which is presumably why Google is comfortable shipping it as fully automatic rather than a manual retouch tool.