Identity Drift and Impossible Motion | Visual AI Learning Guide

Two failures here share one cause. First, identity drift. Compare the first frame with the last. The face is still recognizable, but the features have moved: the nose, the eye spacing, the shape of the jaw. The model is not storing one identity and carrying it forward; it re-creates the subject in each frame from what it has learned faces tend to look like, so small changes accumulate. Second, impossible motion. The model predicts motion that looks plausible rather than simulating forces, so a ball can change direction mid-flight, a limb can pass through a solid object, or hair can move against the body. Watch the clip once at normal speed, then pause at the start and end to compare identity, and trace one moving object to check whether its path is physically possible. Drift plus impossible motion plus flicker is a much stronger signal than any one of them alone.

0:00 / 0:00

Auto-play

Two more video-specific failures come from the same underlying gap: the model has no persistent object and no physics engine.

Identity drift is the gradual change of a face, a logo, or an object across a clip. No single frame looks obviously wrong, but comparing the first frame with the last shows that the person's features have shifted — the nose is slightly different, the eye spacing has changed, a scar has moved, or a jacket's cut has altered. This happens because the model re-generates the subject in each frame from a learned appearance distribution rather than from one stored identity. The drift is usually slow, which is exactly why it is easy to miss at normal playback speed.

Physics inconsistencies come from the model predicting motion that looks plausible rather than simulating forces. Common examples: an object floats or hangs in the air with no support; a thrown ball changes direction mid-flight without a collision; a person's limb passes through a solid object; hair or clothing moves in a direction that contradicts the body's motion; a liquid pours upward or a shadow detaches from its object. The model has learned what motion tends to look like, so the result is often smooth and confident — but it does not obey conservation of momentum or contact constraints.

A practical check is to watch the clip twice: once at normal speed for overall plausibility, and once slowed or paused at the start and end to compare the subject's identity. Then trace one moving object across the clip and ask whether its path and its interactions with other objects are physically possible. Drift and impossible motion are more convincing when they occur together with flicker, because they point to the same cause.