Temporal consistency is the hardest part of video generation because every frame is denoised with its own noise, and nothing forces agreement unless the temporal links carry it. When temporal attention is weak or the clip is long, the links stop holding: the same object is re-decided each frame, so its identity drifts; small per-frame differences in brightness and texture read as flicker; and fast motion outruns the local kernel, producing warping. Independent per-frame generation makes this obvious — each frame is plausible alone, yet the sequence shimmers, because no mechanism ever compared the frames.
How Text-to-Video AI Works Under the Hood