A real model, running on your device
The edge painter is an inpainting network: show it a picture with a gap and it paints what belongs in the gap. Uncropping is that with the gap outside the old frame — the photo sits in a larger canvas, the new border is the gap, and the model continues whatever ran off the edge inward.
It downloads once, from our model file server, and then runs in this tab on WebGPU where your browser offers it and on WebAssembly everywhere else. A second photo reuses the saved model with no download.
Honest about its reach
Within about a third more canvas the result is reliable on open scenes. Past double the size, most of the frame is the model's guess — good as a backdrop, never a claim about what was actually there. The tool shows how much of the result is newly built so you always know.
Questions
Does my photo get uploaded?
No. The picture is read, grown and encoded by code running in this browser tab. It is never sent to a server — there is no server to send it to. The only download that ever happens is the optional edge-painter model, which travels one way, from our model file server to you.
How big is the download?
The edge model and its runtime come to about 40 MB the first time, then they are cached, so later uncrops need no download. The exact size is shown on the button before anything moves.
Does it need a fast GPU?
No. It uses WebGPU when your browser has it and falls back to WebAssembly otherwise. The panel tells you which path ran.
Can AI bring back a person who was cropped out?
No. The model continues the surrounding scene; it cannot recover a specific subject that isn't in the file. Nobody's tool can, because those pixels were discarded at crop time.
What if the result looks wrong?
Switch back to an instant fill, or try a smaller zoom. Busy, detailed edges are harder than plain ones; the honest answer is that the further out you reach, the more the model invents.