How it works
Hold the cube with one corner toward the camera and three faces show: 27 of its 54 stickers. That’s enough, because a slipped move changes about 8 of the 24 stickers you can see.
- 1
Find the cube
A small network looks at the whole frame and answers whether there's a cube, and roughly where its 7 visible corners are.
- 2
Zoom in
A second network looks at a close-up of that spot in the full-quality frame and places the corners exactly. In live video it follows the cube from frame to frame.
- 3
Read the stickers
The corners give a 3 × 3 grid on each of the three faces. A third network reads all 27 stickers together: one of the six colours, or covered by a finger.
- 4
Check the scramble
Each scramble gives the cube you should be holding. The reading is compared with it, trying all 24 ways the cube could be held, and with every version of the scramble with one move slipped. If a slip fits clearly better, it's a mis-scramble.
Private by design
The three networks run in your browser, on your device’s GPU where it has one. Camera frames and scrambles never leave your device, which also keeps competition scrambles secret.
Trained on renders only
The networks learned from tens of thousands of 3D renders of cubes in rooms, in hands, under all kinds of light, with exact labels. On held-out renders the check calls the true scramble correct 99% of the time with no false alarms, and catches 99% of single slipped moves. Results on real photos come next.
CubeCheck helps the scrambler’s own check; it isn’t an official WCA tool. It’s open source under the AGPL, and the models are on Hugging Face.