Plausible Frames Can Make Impossible Motion
A generated clip may look sharp frame by frame while an object accelerates upward under gravity, teleports, disappears, or ignores a declared collision.
physicsvideo evaluates tracked trajectories with explicit tests for acceleration, displacement jumps, visibility gaps, and velocity reversal, then renders the scenarios to real MP4 files.
The Demo
Five scenarios include one clean control and four targeted faults. The benchmark detects all four planted failures and produces zero false positives, for 100% fixture accuracy.
This tests supplied tracks and declared events. General videos require reliable tracking, camera-motion compensation, depth, contact inference, and uncertainty-aware thresholds.
Research Basis
Recent benchmarks treat physical commonsense and temporal consistency as distinct requirements for modern video generation systems.
Read the PhyCoBench.
A Threshold That Graded Its Own Homework
The teleport check compared a track's biggest single-frame jump against a threshold built from that same track's own step sizes — a self-referential design that quietly falls apart under sparse tracking.
Verified directly: a track with only two tracked points — an object seen once, then reappearing 1000 units away one frame later — produced zero violations. Because the adaptive threshold was derived from that single jump's own magnitude, it was mathematically guaranteed to always exceed the jump itself, no matter how extreme the teleport.
Fixed with a leave-one-out threshold: each step's permissiveness limit now comes only from the track's other steps, never from the step being tested. The published demo numbers — four of four planted failures detected, zero false positives, 100% fixture accuracy — are unchanged; only the sparse-track blind spot is closed.