SIFT Β· LoFTR Β· U2-Net β€” 100% local

Stuck on a piece?
Snap it.

Photograph any jigsaw piece and Piece Finder shows you exactly where it belongs on the box art β€” matched two ways by classical and deep computer vision.

Open the app Try the demo ↓
Interactive demo

Tap a piece. Watch it land.

These are real pieces from a real puzzle, located by the actual matching pipeline β€” green is SIFT's projected outline, amber is LoFTR's deep-match region.

Pick a piece
Pick a piece on the left.
How it works

Three snaps to solved.

No account, no training data β€” the models are pretrained and run on the server's CPU.

01

Photograph the puzzle

Box art or the finished picture. It's indexed once β€” thousands of visual landmarks, cached and ready.

02

Photograph one piece

On the table, in your hand, any background β€” a U2-Net segmenter cuts the piece out automatically.

03

See where it goes

Two independent matchers point at the spot, with confidence scores and a zoomable outline on the picture.

Under the hood

Two matchers, because pieces differ.

A sky piece and a castle piece are different problems. Piece Finder runs both matchers side by side and shows every candidate.

SIFT keypoints

Rotation- and scale-invariant features with RANSAC verification β€” a precise, provable outline for textured pieces.

LoFTR transformer

Dense deep matching that still works on sky, water and other low-texture pieces where keypoints starve.

U2-Net segmentation

Cuts the piece from any background before matching, so your photo doesn't need a studio setup.

24/24
real pieces located in evaluation
2Γ—
independent matchers per piece
0
pieces stored β€” piece photos are matched and discarded

Solve the annoying 5%.

The last stretch of every puzzle is a hunt. Make it a photo instead.

Open Piece Finder