When temporal attention links are weak, the model re-decides what each frame contains instead of carrying a decision forward. The result is flicker, where brightness and texture shift frame to frame, and identity drift, where a face or object changes appearance as it moves. The temporal links that should have anchored the same spatial location across frames either never formed or broke, so each frame is solved in isolation. This is the signature failure of insufficient temporal attention, not of conditioning or compression.
How Text-to-Video AI Works Under the Hood