A clown in a storm drain lures a boy with a balloon and pulls him in. A horror shot that tests fast action in three seconds, and a content-filter probe: a child being grabbed is where some models refuse. The prompt leaves the balloon and raincoat colors unstated; a red balloon, a yellow raincoat or a familiar clown face came from the model, not the prompt.
A manual typewriter types RICHMOND, one letter per keystroke. Text in video is where generated letters melt: each letter has to appear when its key strikes, in order, and the word on the paper at the end has to be spelled right.
A girl counts to five on one hand, raising one finger per count, and ends on an open hand. Generated hands are hard in a still and harder in motion: the finger count has to go up by one at a time and the hand has to keep five fingers while it does.
A Scottish fold cat pushes a full glass of water off a kitchen table and it shatters on the tile. Three seconds of cause and effect: the paw has to touch the glass, the glass has to fall at the speed things fall, and glass breaks on tile. The breed is in the prompt because its folded ears are the one detail a model gets wrong by drawing a generic cat.