These text-to-video APIs were tested with one deliberately awkward brief: a small white paper airplane crossing a bright office while the camera tracks it. The test checks more than a pretty first frame. It asks whether a model can keep a small object recognizable, carry it through a moving shot, hold shallow depth of field, and avoid turning the office into a different room halfway through the clip.<\/p>\n
Contents<\/strong><\/p>\n Every run used this prompt: A white paper airplane glides through a sunlit open plan office. Slow tracking shot following the airplane. Shallow depth of field. Realistic.<\/em> The prompt has one moving foreground subject, repeated office geometry, bright window light, and an explicit camera instruction. That makes it useful for spotting subject drift, inconsistent furniture, and camera moves that stop following the action.<\/p>\n This is a one-output comparison, not a benchmark or a ranking. A single generation cannot establish a model-wide quality score. It does show how each selected configuration interprets the same request, and it records the settings and elapsed time from the runs already embedded below.<\/p>\n\n
The text-to-video API test<\/h2>\n
What the five outputs actually show<\/h2>\n
Kling v3<\/h3>\n