This checker verifies whether the objects in an augmented video match the objects the scene is supposed to contain. It answers the question "For every object the World Model expects — a car, a pedestrian, a lane line, a traffic light — did Cosmos render it correctly, per-camera and across cameras?" It matches segmented objects against ground-truth 3D cuboids per camera, grading correspondence and flagging hallucinated objects. By using this checker, we may filter videos for object fidelity, discarding videos whose objects are missing, misplaced, or inconsistent.

License may be found on GitHub: https://github.com/NVIDIA/cosmos-evaluator/blob/main/LICENSE