This service checks whether the motion in an AI-generated video looks physically plausible. It takes a single video and returns a plausibility score and a pass/fail decision, with no reference video or annotations needed. It uses a pretrained video-understanding model, and its decision threshold can be tuned per request. Teams can also retrain or recalibrate it to match their own definition of an acceptable video. It runs on its own or as one of the checkers coordinated by the Arbitrator, which makes it easy to automatically screen large batches of synthetic video for implausible motion and visual artifacts.

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