Self-Refining Video Sampling: inference-time method using a video generator as its own refiner to correct physics and motion. no retraining needed; scores >70% human preference; is validated on Wan2.2 & Cosmos.