Perception does not update frame by frame. The visual system averages input over a few consecutive seconds, remixing what it has already seen with the present moment. This averaging is what keeps the world from flickering, and it is why gradual change inside a clip, like a face slowly drifting, a color creeping, a background reorganizing, is often invisible to the viewer. For generative video, this means that a defect’s visibility depends less on its magnitude than on its rate: what arrives slowly is absorbed into the average, and what arrives abruptly breaks it. Evaluating a model without accounting for this temporal smoothing measures the pixels, not the human percept.