UNTORN

UNTORN reconstructs a torn photo or letter from one overhead image of its separated pieces, then helps a person put the physical original back together. A torn edge can look like a good fit while the writing or image breaks across it, so matching shapes alone is not enough. UNTORN detects the pieces, estimates orientation, proposes rigid edge fits, reads visual continuity across candidate tears, and assembles a page subject to overlap and consistency checks. The result keeps the original photographed pixels, exposes evidence for each accepted seam, and leaves uncertain pieces visible for review. A step-by-step guide gives an order for taping the real pieces back together. Optional fill applies only to enclosed missing-paper regions, is explicitly marked as generated, and never replaces surviving pixels. The interface runs on desktop and mobile, with camera capture, live progress, a reconstruction reveal, before-and-after comparison, seam inspection, text recovery for documents, a local library, and deletion controls. A controlled six-piece letter pile completed the live pipeline with all six pieces placed. Harder photo piles have been checked through offline geometry replay but are not yet claimed as live end-to-end successes. The public evaluation dataset has not been run, so we do not claim a benchmark accuracy.