Every round, you see a shape — its size, position, and rotation — then recreate it from memory. Your score depends on how closely your guess overlaps the original.
Shape starts with pixel-level IoU (Intersection over Union) to compare your guess against the target, curves that overlap score, then applies a symmetry-aware rotation multiplier. Max 10 per round, max 50 over 5 rounds.
Intersection over Union is a standard measure of overlap between two shapes. Both shapes are rasterized onto a canvas, and the algorithm counts how many pixels they share versus the total area either one covers.
This means every dimension matters— position, scale, and rotation all affect overlap. A shape that’s the right size but in the wrong spot scores poorly, and vice versa. The curve gives partial credit for recognizable attempts instead of making every small miss feel fatal.
Shape also applies a separate rotation multiplier because big shapes can overlap a lot even when they’re turned the wrong way. A perfect angle keeps the full score. A badly missed angle can reduce the base score by up to about half on Hard and Brutal.
The memorize timer changes per mode. Shorter exposure time means less time to study the shape before recreating it.
On Easy, targets never rotate — you only need to remember position and scale. On Hard and Brutal, targets can spin up to ±180°, and the allowed position and scale windows widen so shapes can appear anywhere on screen at nearly any size.
Brutalalso adds a 5.5-second input timer. If you don’t submit before it expires, whatever you have on screen is auto-submitted. Very small Brutal targets also get a small scoring boost because tiny shapes are inherently harder to place accurately.
Daily Shape uses the Hard timing and scoring rules, but its target rotations are gentler: from -75° to 75°. Regular Hard, Brutal, async multiplayer, and live multiplayer keep their full rotation range.
These are approximate base scores for overlap alone, before any rotation penalty. A poor angle can pull the final score lower.
We could compute separate deltas for position (Δx, Δy), scale (Δs), and rotation (Δθ) and combine them with weights. But then we’d need to decide: is being 10% off on scale worse than being 10% off on position? Those are arbitrary choices that don’t reflect what the player actually sees.
IoU is still the visual ground truth, but the final score adds just enough rotation awareness to handle cases where overlap alone would be too generous.